Implement revised KA functions `copyto!` and `pagelock!`

[luraess]

JuliaGPU/KernelAbstractions.jl#555

[luraess]

Some additional thoughts from a conversation with @pxl-th which may relate to #799 :

For pagelock!, could one (re-)use unsafe_wrap with lock=false which should pin the memory (https://rocm.docs.amd.com/en/docs-6.0.0/conceptual/gpu-memory.html#pinned-memory)

But, I am unsure about how pagelock!ed memory would interplay with copyto!, if there is a need to dispatch on a different method. My understanding is that pagelock! should act like CUDA.pin which should ensure the host memory to be moved to non-pageable one. Unsure how this interplays in ROCm with mapping though (https://rocm.docs.amd.com/en/docs-6.0.0/conceptual/gpu-memory.html#pinned-memory)

In CUDA, pin mostly calls into register (https://github.com/JuliaGPU/CUDA.jl/blob/6ef1a3dd700448b4902d11e38356bdc58808025f/lib/cudadrv/memory.jl#L167) which page-locks the memory pointed to by ptr. IIRC in AMDGPU, we can achieve pinning by unsafe_wrapping host arrays having (as default) lock=false (https://amdgpu.juliagpu.org/dev/memory/#Wrapping-in-ROCArray) but unsure if one could achieve the CUDA-like "register" step without having to unsafe wrap, i.e., pinning an existing host array.