Harden model weight acquisition and deserialization

[bleugreen]

Summary

• pin the supported DeepRhythm model to an immutable commit URL, filename, byte size, and SHA-256 digest
• verify cached files and stream downloads into same-directory temporary files before atomic publication
• centralize verified acquisition and safe torch.load(..., weights_only=True) deserialization for predictor and batch inference
• add regression coverage for valid and corrupt caches, pinned request parameters, HTTP and stream failures, size and digest rejection, cleanup, atomic replacement, and safe deserialization

Verification

• uv run pytest tests/test_utils.py — 27 passed
• copied weights/deeprhythm-0.7.pth into the cache, then uv run pytest — 68 passed
• uv run ruff check src/ tests/ — passed
• repository search confirms the only torch.load is the shared helper with weights_only=True
• repository search confirms no mutable branch model URL remains