[FEA]: Implement cuda.parallel.copy algorithm

[gevtushenko]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this request and that I agree to the Code of Conduct

Area

Not sure

Is your feature request related to a problem? Please describe.

We need an alternative of thrust::copy_n algorithm in cuda.parallel module.

Describe the solution you'd like

Asynchronous cuda.parallel.copy_n algorithm that accepts host, device, device, host, or device, device iterators and performs a copy in-place (no temporaries).

Describe alternatives you've considered

No response

Additional context

No response

[shwina]

One thing that came up in a sync with @gevtushenko today is that we would ideally like for copy (and friends) to have a very simple API, for example:

copy(from=d_input, to=h_output)
copy(from=d_input, to=d_output)

as opposed to:

copy(from=d_input, to=h_output, how="device_to_host")
copy(from=d_input, to=d_output, how="device_to_device")

Whereas in C++ the argument types can be used to dispatch to the appropriate CUDA API, we must rely on less formal protocols in Python (like __cuda_array_interface__) to do the same.