[FEA]: Implement <cuda/experimental/file>

[davebayer]

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

CUDA Experimental (cudax)

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

I've been wondering whether it would be a good idea to add a wrapper of cuFile to cudax.

Describe the solution you'd like

Implement <cuda/experimental/file> module with:

• cudax::file class wrapping operations with CUfileHandle_t
• cudax::file_driver object implementing cuFileDriver*() operations
• cudax::file_buffer_registerer RAII wrapper class for registering buffers in cuFile driver
• cudax::file_stream_registerer RAII wrapper class for registering stream in cuFile driver
• some more classes for batched IO

I've tried to implement it in #3360 without the batched IO. It is just a not tested proposal, may contain errors.

What do you guys think?

Describe alternatives you've considered

No response

Additional context

No response

[leofang]

I think there's already prior art and it's maintained by the RAPIDS (& cuFILE) teams: https://github.com/rapidsai/kvikio. There's already a nice C++ abstraction over cuFILE (and beyond): https://github.com/rapidsai/kvikio/tree/branch-25.02/cpp/include/kvikio. I'd suggest to take inputs from the RAPIDS-kvikio team.

cc @madsbk @jakirkham for vis

[jakirkham]

cc @akshaysubr (for awareness)