Type unstable code can lead to loading from CPU boxes

**Describe the bug**

In some cases assigning NaN to a cuda array of Float32 results in:

CUDA error: an illegal memory access was encountered (code 700, ERROR_ILLEGAL_ADDRESS)

assigning a NaN32 does not.

**To reproduce**

The Minimal Working Example (MWE) for this bug:

```julia
using CUDA
CUDA.allowscalar(false)

#return NaN if sum of colum c of A is 0, 0 otherwise
function core_test(A,c)
n = Int32(0)
for i in 1:size(A,1)
n += A[i,c]
end
if n == 0
return NaN # <===== NaN32 doesnt crash
else
return 0.0f0
end
end

#run core_test on each colum of A
function test_func!(out,A)
index = (blockIdx().x -1) * blockDim().x + threadIdx().x
stride = blockDim().x * gridDim().x
for c = index:stride:size(out,1)
@inbounds out[c] = core_test(A,c)
end
return nothing
end

A = cu([ 1 0 ])
out = CUDA.fill(NaN32, 2)

CUDA.@sync begin
@cuda test_func!(out,A)
end
@show out

```

Manifest.toml

```
CUDA: v2.1.0
GPUArrays: V6.1.1

[deps]
Adapt = "79e6a3ab-5dfb-504d-930d-738a2a938a0e"
Arrow = "69666777-d1a9-59fb-9406-91d4454c9d45"
BenchmarkTools = "6e4b80f9-dd63-53aa-95a3-0cdb28fa8baf"
CSV = "336ed68f-0bac-5ca0-87d4-7b16caf5d00b"
CUDA = "052768ef-5323-5732-b1bb-66c8b64840ba"
DataFrames = "a93c6f00-e57d-5684-b7b6-d8193f3e46c0"
GPUArrays = "0c68f7d7-f131-5f86-a1c3-88cf8149b2d7"
IJulia = "7073ff75-c697-5162-941a-fcdaad2a7d2a"
Missings = "e1d29d7a-bbdc-5cf2-9ac0-f12de2c33e28"
OnlineStats = "a15396b6-48d5-5d58-9928-6d29437db91e"
Plots = "91a5bcdd-55d7-5caf-9e0b-520d859cae80"
Pluto = "c3e4b0f8-55cb-11ea-2926-15256bba5781"
Query = "1a8c2f83-1ff3-5112-b086-8aa67b057ba1"
RollingFunctions = "b0e4dd01-7b14-53d8-9b45-175a3e362653"
ShiftedArrays = "1277b4bf-5013-50f5-be3d-901d8477a67a"
StaticArrays = "90137ffa-7385-5640-81b9-e52037218182"
StatsBase = "2913bbd2-ae8a-5f71-8c99-4fb6c76f3a91"
StatsPlots = "f3b207a7-027a-5e70-b257-86293d7955fd"
ThreadTools = "dbf13d8f-d36e-4350-8970-f3a99faba1a8"
Transducers = "28d57a85-8fef-5791-bfe6-a80928e7c999"
```

**Expected behavior**

I'd expect the same behavior as assign NaN32 or a compile-time error.

**Version info**

Details on Julia:

```
# please post the output of:
versioninfo()

Julia Version 1.5.2
Commit 539f3ce943 (2020-09-23 23:17 UTC)
Platform Info:
OS: Linux (x86_64-pc-linux-gnu)
CPU: AMD Ryzen 7 1700 Eight-Core Processor
WORD_SIZE: 64
LIBM: libopenlibm
LLVM: libLLVM-9.0.1 (ORCJIT, znver1)
Environment:
JULIA_NUM_THREADS = 15
JULIA_EDITOR = "/usr/share/code/code"
```

Details on CUDA:

```
# please post the output of:
CUDA.versioninfo()

UDA toolkit 11.0.3, artifact installation
CUDA driver 11.0.0
NVIDIA driver 450.80.2

Libraries:
- CUBLAS: 11.2.0
- CURAND: 10.2.1
- CUFFT: 10.2.1
- CUSOLVER: 10.6.0
- CUSPARSE: 11.1.1
- CUPTI: 13.0.0
- NVML: 11.0.0+450.80.2
- CUDNN: 8.0.4 (for CUDA 11.0.0)
- CUTENSOR: 1.2.1 (for CUDA 11.0.0)

Toolchain:
- Julia: 1.5.2
- LLVM: 9.0.1
- PTX ISA support: 3.2, 4.0, 4.1, 4.2, 4.3, 5.0, 6.0, 6.1, 6.3, 6.4
- Device support: sm_35, sm_37, sm_50, sm_52, sm_53, sm_60, sm_61, sm_62, sm_70, sm_72, sm_75

2 devices:
0: GeForce GTX 1060 6GB (sm_61, 5.784 GiB / 5.934 GiB available)
1: GeForce GTX 1060 3GB (sm_61, 93.438 MiB / 2.949 GiB available)

```

**Additional context**

Add any other context about the problem here.