Expose cuFile stream register API to Python

[kingcrimsontianyu]

This PR exposes cuFile stream register and deregister functions to KvikIO Python API.

kvikio.stream_register(stream, flags)
kvikio.stream_deregister(stream)

The stream register and deregister is an optional, performance-tuning API. The users specify the precondition of the I/O parameters by setting the flags that indicate when the I/O parameters become valid and whether they are page aligned.

The added Python API can be unit tested using:

pytest ~/kvikio/python/kvikio/tests/test_async_io.py::test_stream_register_deregister -m cufile -vsx

A known issue is that, under CUDA 12.9, which is the only setting tested, if the file path is not on a drive with full GDS support, the above unit test will pass but end with a segmentation fault. This is beyond the scope of this PR and is likely not an issue in KvikIO. Further investigation is a good-to-have.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[madsbk]

We might consider using something like the following for the CUDA Stream bindings

from rmm.pylibrmm.stream import Stream

https://docs.rapids.ai/api/rmm/stable/python/pylibrmm/#rmm.pylibrmm.stream.Stream

[kingcrimsontianyu]

I've changed the parameter name to raw_stream for the Python API. For your suggestion, do you think we need to add Python RMM dependency to KvikIO, or add KvikIO's own stream wrapper on C++ and Python level? @madsbk

[madsbk]

I like the raw_stream name.

For your suggestion, do you think we need to add Python RMM dependency to KvikIO, or add KvikIO's own stream wrapper on C++ and Python level? @madsbk

At some point, yes. Or use cuda-python's CUDA stream, or roll our own :)

[kingcrimsontianyu]

That makes sense. I've created this ticket #895 .
Re-requested the review after a minor update to the cufile.py which uses the raw stream as well.