feat: optional message embedding for semantic search (config-driven)

Adds optional message embedding support to Hermes Agent's memory system, enabling hybrid search (FTS5 full-text + vector cosine similarity) for more relevant memory retrieval.

What it does

• Schema v12: adds message_embedding BLOB column to messages table (auto-migrated on startup)
• Config-driven: embedding is disabled by default; enable by adding to ~/.hermes/config.yaml:

  embedding:
    base_url: http://your-embedding-server:8081   # Ollama-compatible /v1/embeddings
    model: Qwen3-Embedding-0.6B
    dimension: 1024

• Graceful fallback: if no config, pure FTS5 is used (zero overhead)
• Hybrid search: FTS5 results are re-ranked by cosine similarity using stored embeddings
• Vector-only fallback: if FTS5 returns nothing, falls back to pure vector search
• Backfill script: scripts/backfill_embeddings.py computes embeddings for existing messages

Files changed

• hermes_state.py — core embedding logic + schema migration
• scripts/backfill_embeddings.py — utility to backfill embeddings for existing DB rows

Design notes

• Embeddings are computed via Ollama-compatible /v1/embeddings API (tested with Qwen3-Embedding-0.6B on a 2080ti)
• Vectors stored as packed float32 (struct.pack('f' * dim, ...)) in SQLite BLOB
• Schema migration is non-destructive: existing rows get NULL embedding, filled lazily or via backfill script
• No external dependencies added — uses urllib.request (standard library) for the embedding API call

Testing

• Schema migration (v11→v12) tested manually
• Hybrid search tested with a local embedding server
• Graceful fallback (no config) verified — pure FTS5, no errors
• Backfill script tested with --dry-run and --apply modes