Publish CUDA wheel variants on Hugging Face

[andimarafioti]

Summary

Add a Hugging Face wheel publishing path for CUDA-specific qwentts-cpp-python variants while keeping PyPI as the CUDA 12.8 default.

This PR adds:

• CUDA 12.4 wheel validation artifacts in the existing PR wheel workflow.
• A new Publish Hugging Face Wheels workflow that builds local-version CUDA wheels:
  • 0.2.0+cu124
  • 0.2.0+cu128
  • 0.2.0+cu130
• Static --find-links HTML pages for each CUDA flavor.
• An upload script that creates the public HF dataset repo if needed and uploads the prepared wheel tree using the HF_TOKEN repository secret.
• README install instructions for PyPI and HF-hosted CUDA variants.

Why --find-links

PyTorch uses custom package indexes such as https://download.pytorch.org/whl/cu128. A plain Hugging Face repo is raw file hosting, not a PEP 503 directory index, so the reliable equivalent is a static HTML wheel page consumed by pip via --find-links:

pip install "qwentts-cpp-python==0.2.0+cu130" \
  -f https://huggingface.co/datasets/andimarafioti/qwentts-cpp-python-wheels/resolve/main/whl/cu130.html

Dependencies still resolve from PyPI normally.

CUDA matrix

• CUDA 12.4: Ubuntu 22.04 CUDA repo, native sm_80, sm_86, sm_90, PTX sm_75.
• CUDA 12.8: Ubuntu 24.04 CUDA repo, native sm_80, sm_86, sm_90, sm_120, PTX sm_75, sm_120 for HF wheels.
• CUDA 13.0: Ubuntu 24.04 CUDA repo, native sm_80, sm_86, sm_90, sm_121, PTX sm_75.

The existing PyPI publish workflow remains unchanged and still publishes CUDA 12.8 wheels only.

Validation

• Parsed all workflow YAML files with PyYAML.
• Compiled the new helper scripts with py_compile.
• Tested local-version rewriting on a temporary repo copy.
• Tested HF wheel-index generation with temporary wheel filenames.
• Confirmed CUDA 12.4 packages are available in NVIDIA's Ubuntu 22.04 x86_64 and sbsa repos, but not Ubuntu 24.04.
• Ran package tests with uv run --with pytest pytest -q tests: 5 passed, 1 skipped.

Setup Needed

Add a repository secret named HF_TOKEN with write access to the target HF account/org. The workflow will create the public dataset repo andimarafioti/qwentts-cpp-python-wheels if it does not exist.