Handling wrapper array inputs

[gaurav-arya]

We do not correctly handle wrapper array input types for non-CPU arrays. For example:

julia> using CUDA.CUFFT

julia> P = plan_fft(CuArray([1,2,3]))
CUFFT d.p. complex forward plan for 3-element CuArray of ComplexF64

julia> x = CuArray([1.0 2.0; 3.0 4.0])
2×2 CuArray{Float64, 2, CUDA.Mem.DeviceBuffer}:
 1.0  2.0
 3.0  4.0

julia> P = plan_fft(CuArray([1,2,3]))
CUFFT d.p. complex forward plan for 3-element CuArray of ComplexF64

julia> P * view(x, axes(x)...)
ERROR: StackOverflowError:
Stacktrace:
 [1] Array
   @ AbstractFFTs ./boot.jl:477 [inlined]
 [2] Array
   @ AbstractFFTs ./boot.jl:485 [inlined]
 [3] Array
   @ AbstractFFTs ./boot.jl:492 [inlined]
 [4] copy1(::Type{ComplexF64}, x::Vector{ComplexF64})
   @ AbstractFFTs ~/AbstractFFTs.jl/src/definitions.jl:53 [inlined]
 [5] *(p::CUDA.CUFFT.cCuFFTPlan{ComplexF64, -1, false, 1}, x::Matrix{ComplexF64}) (repeats 64932 times)
   @ AbstractFFTs ~/AbstractFFTs.jl/src/definitions.jl:224
 [6] *(p::CUDA.CUFFT.cCuFFTPlan{ComplexF64, -1, false, 1}, x::SubArray{Float64, 2, CuArray{…}, Tuple{…}, false})
   @ AbstractFFTs ~/AbstractFFTs.jl/src/definitions.jl:224
 [7] top-level scope
   @ REPL[80]:1
Some type information was truncated. Use `show(err)` to see complete types.

I don't know whether there is a standard solution to this issue? Possibly we would want to design a cleverer fallback * dispatch, that, perhaps using parent, would find its way to the more specific dispatch.

Related: #32, #73