Inconsistent `KernelError` on the GPU telling me to use `return` in my kernel

Was trying to debug an issue with the use of `^` and `pow` inside `@kernel`s when I came across this weird (and incorrect/inconsistent) warning.

```julia
using KernelAbstractions
using CUDAKernels
using CUDA

@kernel function apply_f_kernel!(A, B, f)
I = @index(Global)
@inbounds A[I] = f(B[I])
end

N = 3

square(x) = x^2

for f in [sin, cos, square]
A = zeros(N)
B = zeros(N)

@info "Calling apply_f_kernel! on the CPU with f=$f..."
event = apply_f_kernel!(CPU(), N)(A, B, f, ndrange=N)
wait(event)

A = zeros(N) |> CuArray
B = zeros(N) |> CuArray

@info "Calling apply_f_kernel! on the GPU with f=$f..."
event = apply_f_kernel!(CUDADevice(), N)(A, B, ndrange=N)
wait(event)
end
```

produces

```
[ Info: Calling apply_f_kernel! on the CPU with f=sin...
[ Info: Calling apply_f_kernel! on the GPU with f=sin...
ERROR: LoadError: GPU compilation of kernel gpu_apply_f_kernel!(Cassette.Context{nametype(CUDACtx), KernelAbstractions.CompilerMetadata{KernelAbstractions.NDIteration.DynamicSize, KernelAbstractions.NDIteration.DynamicCheck, Nothing, CartesianIndices{1, Tuple{Base.OneTo{Int64}}}, KernelAbstractions.NDIteration.NDRange{1, KernelAbstractions.NDIteration.DynamicSize, KernelAbstractions.NDIteration.StaticSize{(3,)}, CartesianIndices{1, Tuple{Base.OneTo{Int64}}}, Nothing}}, Nothing, KernelAbstractions.var"##PassType#257", Nothing, Cassette.DisableHooks}, typeof(gpu_apply_f_kernel!), CuDeviceVector{Float64, 1}, CuDeviceVector{Float64, 1}) failed
KernelError: kernel returns a value of type `Union{}`

Make sure your kernel function ends in `return`, `return nothing` or `nothing`.
If the returned value is of type `Union{}`, your Julia code probably throws an exception.
Inspect the code with `@device_code_warntype` for more details.

Stacktrace:
[1] check_method(job::GPUCompiler.CompilerJob)
@ GPUCompiler ~/.julia/packages/GPUCompiler/eJOtJ/src/validation.jl:21
[2] macro expansion
@ ~/.julia/packages/TimerOutputs/4QAIk/src/TimerOutput.jl:206 [inlined]
[3] macro expansion
@ ~/.julia/packages/GPUCompiler/eJOtJ/src/driver.jl:88 [inlined]
[4] emit_julia(job::GPUCompiler.CompilerJob)
@ GPUCompiler ~/.julia/packages/GPUCompiler/eJOtJ/src/utils.jl:62
[5] cufunction_compile(job::GPUCompiler.CompilerJob)
@ CUDA ~/.julia/packages/CUDA/LTbUr/src/compiler/execution.jl:299
[6] check_cache
@ ~/.julia/packages/GPUCompiler/eJOtJ/src/cache.jl:47 [inlined]
[7] cached_compilation
@ ~/.julia/packages/Cassette/FwMN0/src/overdub.jl:582 [inlined]
[8] cufunction(f::typeof(Cassette.overdub), tt::Type{Tuple{Cassette.Context{nametype(CUDACtx), KernelAbstractions.CompilerMetadata{KernelAbstractions.NDIteration.DynamicSize, KernelAbstractions.NDIteration.DynamicCheck, Nothing, CartesianIndices{1, Tuple{Base.OneTo{Int64}}}, KernelAbstractions.NDIteration.NDRange{1, KernelAbstractions.NDIteration.DynamicSize, KernelAbstractions.NDIteration.StaticSize{(3,)}, CartesianIndices{1, Tuple{Base.OneTo{Int64}}}, Nothing}}, Nothing, KernelAbstractions.var"##PassType#257", Nothing, Cassette.DisableHooks}, typeof(gpu_apply_f_kernel!), CuDeviceVector{Float64, 1}, CuDeviceVector{Float64, 1}}}; name::String, kwargs::Base.Iterators.Pairs{Union{}, Union{}, Tuple{}, NamedTuple{(), Tuple{}}})
@ CUDA ~/.julia/packages/CUDA/LTbUr/src/compiler/execution.jl:289
[9] macro expansion
@ ~/.julia/packages/CUDA/LTbUr/src/compiler/execution.jl:102 [inlined]
[10] (::KernelAbstractions.Kernel{CUDADevice, KernelAbstractions.NDIteration.StaticSize{(3,)}, KernelAbstractions.NDIteration.DynamicSize, typeof(gpu_apply_f_kernel!)})(::CuArray{Float64, 1}, ::Vararg{CuArray{Float64, 1}, N} where N; ndrange::Int64, dependencies::CUDAKernels.CudaEvent, workgroupsize::Nothing, progress::Function)
@ CUDAKernels ~/.julia/packages/CUDAKernels/cz35Z/src/CUDAKernels.jl:194
[11] top-level scope
@ ~/Oceananigans.jl/apply_f_error.jl:26
[12] include(fname::String)
@ Base.MainInclude ./client.jl:444
[13] top-level scope
@ REPL[14]:1
[14] top-level scope
@ ~/.julia/packages/CUDA/LTbUr/src/initialization.jl:81
in expression starting at /home/alir/Oceananigans.jl/apply_f_error.jl:14
```

which tells me to use `return` in the kernel but KernelAbstractions.jl doesn't allow that.

# Environment

```
julia> versioninfo()
Julia Version 1.6.1
Commit 6aaedecc44 (2021-04-23 05:59 UTC)
Platform Info:
OS: Linux (x86_64-pc-linux-gnu)
CPU: Intel(R) Xeon(R) Silver 4214 CPU @ 2.20GHz
WORD_SIZE: 64
LIBM: libopenlibm
LLVM: libLLVM-11.0.1 (ORCJIT, cascadelake)
```

Manifest

```
(Oceananigans) pkg> status -m
Project Oceananigans v0.57.1
Status `~/Oceananigans.jl/Manifest.toml`
[621f4979] AbstractFFTs v1.0.1
[79e6a3ab] Adapt v3.3.0
[4fba245c] ArrayInterface v3.1.11
[ab4f0b2a] BFloat16s v0.1.0
[fa961155] CEnum v0.4.1
[179af706] CFTime v0.1.1
[052768ef] CUDA v3.2.0
[72cfdca4] CUDAKernels v0.2.0
[7057c7e9] Cassette v0.3.6
[d360d2e6] ChainRulesCore v0.9.42
[34da2185] Compat v3.28.0
[a8cc5b0e] Crayons v4.0.4
[7445602f] CubedSphere v0.1.0
[9a962f9c] DataAPI v1.6.0
[864edb3b] DataStructures v0.18.9
[e2d170a0] DataValueInterfaces v1.0.0
[b552c78f] DiffRules v1.0.2
[ffbed154] DocStringExtensions v0.8.4
[b305315f] Elliptic v1.0.1
[e2ba6199] ExprTools v0.1.3
[7a1cc6ca] FFTW v1.4.1
[5789e2e9] FileIO v1.8.2
[0c68f7d7] GPUArrays v6.4.0
[61eb1bfa] GPUCompiler v0.11.5
[c27321d9] Glob v1.3.0
[615f187c] IfElse v0.1.0
[82899510] IteratorInterfaceExtensions v1.0.0
[033835bb] JLD2 v0.4.5
[692b3bcd] JLLWrappers v1.3.0
[0f8b85d8] JSON3 v1.8.1
[63c18a36] KernelAbstractions v0.6.2
[929cbde3] LLVM v3.7.0
[da04e1cc] MPI v0.17.2
[1914dd2f] MacroTools v0.5.6
[c03570c3] Memoize v0.4.4
[85f8d34a] NCDatasets v0.11.4
[77ba4419] NaNMath v0.3.5
[6fe1bfb0] OffsetArrays v1.8.0
[bac558e1] OrderedCollections v1.4.1
[69de0a69] Parsers v1.1.0
[0e08944d] PencilArrays v0.9.4
[4a48f351] PencilFFTs v0.12.2
[21216c6a] Preferences v1.2.2
[74087812] Random123 v1.3.1
[e6cf234a] RandomNumbers v1.4.0
[189a3867] Reexport v1.0.0
[ae029012] Requires v1.1.3
[6038ab10] Rotations v1.0.2
[1bc83da4] SafeTestsets v0.0.1
[6c6a2e73] Scratch v1.0.3
[d496a93d] SeawaterPolynomials v0.2.0
[276daf66] SpecialFunctions v1.3.0
[aedffcd0] Static v0.2.4
[90137ffa] StaticArrays v1.1.3
[15972242] StaticPermutations v0.3.0
[09ab397b] StructArrays v0.5.1
[856f2bd8] StructTypes v1.7.2
[3783bdb8] TableTraits v1.0.1
[bd369af6] Tables v1.4.2
[6aa5eb33] TaylorSeries v0.10.13
[a759f4b9] TimerOutputs v0.5.8
[3bb67fe8] TranscodingStreams v0.9.5
[bc48ee85] Tullio v0.2.14
[f5851436] FFTW_jll v3.3.9+7
[0234f1f7] HDF5_jll v1.12.0+1
[1d5cc7b8] IntelOpenMP_jll v2018.0.3+2
[856f044c] MKL_jll v2021.1.1+1
[7cb0a576] MPICH_jll v3.3.2+10
[9237b28f] MicrosoftMPI_jll v10.1.3+0
[7243133f] NetCDF_jll v400.702.400+0
[fe0851c0] OpenMPI_jll v4.0.2+2
[458c3c95] OpenSSL_jll v1.1.1+6
[efe28fd5] OpenSpecFun_jll v0.5.4+0
[0dad84c5] ArgTools
[56f22d72] Artifacts
[2a0f44e3] Base64
[ade2ca70] Dates
[8bb1440f] DelimitedFiles
[8ba89e20] Distributed
[f43a241f] Downloads
[b77e0a4c] InteractiveUtils
[4af54fe1] LazyArtifacts
[b27032c2] LibCURL
[76f85450] LibGit2
[8f399da3] Libdl
[37e2e46d] LinearAlgebra
[56ddb016] Logging
[d6f4376e] Markdown
[a63ad114] Mmap
[ca575930] NetworkOptions
[44cfe95a] Pkg
[de0858da] Printf
[3fa0cd96] REPL
[9a3f8284] Random
[ea8e919c] SHA
[9e88b42a] Serialization
[1a1011a3] SharedArrays
[6462fe0b] Sockets
[2f01184e] SparseArrays
[10745b16] Statistics
[fa267f1f] TOML
[a4e569a6] Tar
[8dfed614] Test
[cf7118a7] UUIDs
[4ec0a83e] Unicode
[e66e0078] CompilerSupportLibraries_jll
[deac9b47] LibCURL_jll
[29816b5a] LibSSH2_jll
[c8ffd9c3] MbedTLS_jll
[14a3606d] MozillaCACerts_jll
[83775a58] Zlib_jll
[8e850ede] nghttp2_jll
[3f19e933] p7zip_jll
```

Start with the reproducer in apply_f_error.jl and the @kernel apply_f_kernel! entry point, then trace the GPU compilation path through CUDAKernels.jl:194 and GPUCompiler's validation.jl:21. Done means the GPU diagnostic is consistent with KernelAbstractions kernel rules and no longer gives incorrect return guidance for this case.

Written by the indexing model from the issue text.