refactor(embedding_compute): extract providers into adapter pattern

[luojiyin1987]

Summary

Split the 1444-line embedding_compute.py monolith into a registry-backed adapter pattern.

Changes

• embedding_compute.py: 1444 → 403 lines — keeps shared utilities + compute_embeddings() dispatch + register_provider() API
• New providers/ package — one module per backend:
  • sentence_transformers.py (487 lines)
  • ollama.py (335 lines)
  • openai.py (136 lines)
  • gemini.py (89 lines)
  • mlx.py (82 lines)

Design

compute_embeddings(texts, model_name, mode="sentence-transformers")
    │
    ├─ _init_providers()  → lazy-import all providers
    ├─ _providers[mode]   → dispatch via registry dict
    └─ register_provider("new-mode", fn)  → third-party extensibility

Benefits

• Adding a new provider: 1 new file + 1 registry line (was: touch 1444-line elif chain)
• Adding cross-encoder reranker: just register_provider("reranker", fn)
• Full backward compat — all ~90 callers unchanged
• Each provider independently testable