Add sm_90a to enable use of accelerated wgmma and setmaxnreg instructions

[ConsceIeratus]

sm_90a adds support for accelerated wgmma and setmaxnreg instructions.

https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#ptx-module-directives-target

[TimDettmers]

Thank you, this is good to know that wgmma is now added. I think Hopper supports both sm_90 and sm_90a. Since we do not make use of wgmma or setmaxnreg for now we would not need sm_90a. I would want to not add it at the moment to keep the binary a bit smaller. I am currently having troubles with the binary size since all binaries must be smaller than 100MB for PyPi uploads.

[Titus-von-Koeller]

Thanks for raising this. We'll keep this in mind and implement it once we have figured out the current cross-platform + build + distribution topics.

[akx]

For PyPI, you can request a quota increase for the project via https://github.com/pypi/support :)

[younesbelkada]

Nice - @akx would you mind sharing that in https://github.com/TimDettmers/bitsandbytes/discussions/990 as well 🙏

[Titus-von-Koeller]

For PyPI, you can request a quota increase for the project via https://github.com/pypi/support :)

Nice, that would unblock us on this before we've figured out the cross-platform compilation and distribution stuff.

I'll look into that today. Valuable input, thanks!