Fix padding causing device to host copies

[pxl-th]

Current version of pad_reflect causes input array to be copied to host during backpropagation because of this rule.

Instead of usign StepRange we should use UnitRange for selectdim and reverse, this avoids expensive copying.

julia> using Revise, CUDA, Zygote

julia> x = CuArray(rand(Float32, 8000, 1));

julia> Δ = CuArray(ones(Float32, 1, 1));

julia> y, back = Zygote.pullback(x) do x
           sum(NNlib.pad_reflect(x, (0, 10)); dims=(1, 2))
       end;

julia> back(Δ);
D to H copy CuArray{Float32, 2, CUDA.DeviceMemory}: (8000, 1) -> Matrix{Float32}: (8000, 1)
D to H copy CuArray{Float32, 2, CUDA.DeviceMemory}: (11, 1) -> Matrix{Float32}: (11, 1)

Besides copying needlessly, on AMDGPU device to host copy causes synchronization if the memory is unpinned or there is no hardware support for this.

Before:

julia> @btime AMDGPU.@sync back($Δ);
  233.065 μs (330 allocations: 75.60 KiB)

Now:

julia> @btime AMDGPU.@sync back($Δ);
  80.891 μs (293 allocations: 11.63 KiB)

Same for pad_symmetric:

Before:

julia> @btime AMDGPU.@sync back($Δ);
  243.315 μs (327 allocations: 75.55 KiB)

Now:

julia> @btime AMDGPU.@sync back($Δ);
  75.781 μs (293 allocations: 11.62 KiB)

PR Checklist

• Tests are added
• Documentation, if applicable