CI: HuggingFace cache + offline runs for feature_utils tests

[lmeyerov]

Summary

Feature utils tests hit HF rate limits (429) during model downloads. Add a deterministic HF cache + offline mode in CI.

Proposal

• Add a warmup + cache job before matrix runs:
  • Set HF_HOME (or HF_HUB_CACHE) to ${{ runner.temp }}/hf-cache.
  • Use actions/cache with key hf-cache-${{ runner.os }}-v1 (stable per-OS; no per-Python split).
  • Warm once using HF_TOKEN secret for:
    • sentence-transformers/average_word_embeddings_komninos
    • sentence-transformers/paraphrase-MiniLM-L6-v2
    • sentence-transformers/paraphrase-albert-small-v2
  • After restore/warm, set HF_HUB_OFFLINE=1 for all test steps.
• All matrix jobs restore the same cache; bump key version only when model set changes.
• Optional: pre-bake the cache into GHCR or a release asset for zero-network runs.
• Optional: lower parallelism for HF-heavy jobs if needed.

Rationale

• Unblocks graphistry/tests/test_feature_utils.py failures (429s/missing HF files).
• Makes CI deterministic and faster by avoiding external throttling.

Notes

• Models are ~250MB total, fine for cache/asset.
• If HF_TOKEN is missing, skip warm and rely on existing cache or skip HF tests.