Add CPU wheels to Hugging Face variants

[andimarafioti]

Summary

• add CPU x86_64 and aarch64 wheel jobs to the Hugging Face wheel publisher
• publish those CPU wheels as the 0.2.0+cpu local-version flavor alongside CUDA flavors
• update README, AGENTS.md, and generated HF README wording from CUDA-only to backend-specific

Validation

• python3 -c 'import yaml; [yaml.safe_load(open(p)) for p in [\".github/workflows/publish-hf-wheels.yml\", \".github/workflows/wheels.yml\", \".github/workflows/publish.yml\"]]; print(\"workflow yaml ok\")'
• python3 -m py_compile scripts/prepare_hf_wheel_repo.py
• uv run --with pytest pytest -q tests
• smoke-tested scripts/prepare_hf_wheel_repo.py with fake +cpu and +cu130 wheels; it generated whl/cpu.html, whl/cu130.html, and whl/index.html