[CI] Generate CUDA SBSA Wheels

[johnnynunez]

Summary

Add CI support for building CUDA Python wheels on ARM64 (SBSA) hosts, complementing the existing x86_64 CUDA wheel pipeline.

• New workflow (.github/workflows/ubuntu-wheel-sbsa.yml): Builds and tests CUDA + CPU wheels on ubuntu-24.04-arm runners for Python 3.10–3.14, using the same docker/docker_build.sh cuda_wheel_py3XX[_dev] commands as x86.
• Dockerfile.wheel: Make the NVIDIA apt-key rotation fix conditional on x86_64 — ARM CUDA images are unaffected and the old ubuntu1604/x86_64 key URLs don't exist for aarch64.
• docker_build.sh: cuda_wheel_build() now selects an architecture-specific ccache tar name (open3d-ubuntu-2204-cuda-sbsa-ci-ccache on aarch64 vs open3d-ubuntu-2204-cuda-ci-ccache on x86_64) so the two caches never collide.

Usage

Locally on an ARM64 machine (or on the ubuntu-24.04-arm runner):

# Developer build
docker/docker_build.sh cuda_wheel_py312_dev

# Release build
docker/docker_build.sh cuda_wheel_py312

The same commands work on x86_64 — architecture is auto-detected via uname -m.

Test plan

• Verify the Ubuntu Wheel SBSA workflow triggers on this PR and builds the Python 3.14 wheel (the only version not excluded for PRs)
• Verify the existing Ubuntu Wheel (x86) workflow is unaffected and still passes
• Confirm produced .whl artifacts contain manylinux + aarch64 in the filename
• On main, verify all Python versions (3.10–3.14) are built and uploaded to the devel release

More information:
https://developer.nvidia.com/blog/cuda-13-2-introduces-enhanced-cuda-tile-support-and-new-python-features/

[update-docs]

Thanks for submitting this pull request! The maintainers of this repository would appreciate if you could update the CHANGELOG.md based on your changes.