Recurse into some arrays of arrays?

I wonder whether `adapt` should treat an Array of Arrays as a container, like a Tuple, rather than storage to be converted. Convert the innermost Array, not the outermost:
```julia
julia> (rand(1,1), rand(1,1)) |> cu |> first
1×1 CuArray{Float32, 2, CUDA.Mem.DeviceBuffer}:
0.37755206

julia> [rand(1,1), rand(1,1)] |> cu # should perhaps work as above
ERROR: CuArray only supports element types that are allocated inline.
Matrix{Float64} is a mutable type

julia> [1:2, 3:4] |> cu # should not change
2-element CuArray{UnitRange{Int64}, 1, CUDA.Mem.DeviceBuffer}:
1:2
3:4

julia> [SA[1,2.], SA[3,4.]] |> cu
2-element CuArray{SVector{2, Float64}, 1, CUDA.Mem.DeviceBuffer}:
[1.0, 2.0]
[3.0, 4.0]

julia> Any[1:2, 3:4] |> cu
ERROR: CuArray only supports element types that are allocated inline.
```
I'm not exactly sure what the rule would be, perhaps something like `isbitstype(eltype(x))` is enough?

This came up in https://github.com/JuliaGPU/CUDA.jl/pull/1769, where CuIterator at present produces a `Vector{CuArray}`.

The issue centers on Adapt.adapt for nested Julia Arrays and references CUDA.jl's CuIterator; start by tracing that adaptation entry point and compare the tuple, typed-range, StaticArrays, and Any examples. Define the supported recursion rule and verify each example's resulting behavior, including the failure cases.

Written by the indexing model from the issue text.