ARM runners (faster than cross compilation qemu)

[johnnynunez]

This pull request includes updates to the .github/workflows/python-package.yml file to expand the testing matrix and update a GitHub action version. The most important changes include adding a new operating system to the matrix and updating the version of the softprops/action-gh-release action.

Updates to testing matrix:

• Added ubuntu-22.04-arm to the os matrix for various jobs, expanding the range of environments tested. [1] [2] [3]
• Updated the exclusion rules to account for the new ubuntu-22.04-arm environment. [1] [2] [3]

GitHub action update:

• Updated softprops/action-gh-release from version v2.0.8 to v2.2.1 to ensure compatibility and access to new features.

cc @matthewdouglas

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[johnnynunez]

@matthewdouglas i downgrade to 22.04 for glibc. Also affects to cuda builds (I tested on jetson)

[matthewdouglas]

Related to #1437

[johnnynunez]

Related to #1437

we distribute for jetson and sbsa: #1437 (comment)

[matthewdouglas]

Hi @johnnynunez, thanks for your patience here! I tried out an aarch64 sbsa build on an AWS g5g instance (T4/sm75). The unit tests we have for GPU and CPU all passed as expected.

I've resolved some conflicts here and made a couple of small changes.

I'm thinking we should only target sm75, sm80, sm90, and sm100. Thoughts on that?

[matthewdouglas]

I'm going to merge and handle the change to limit aarch64 CUDA architectures in a separate PR.