Speed up JSONL hydration

[valiantone]

What

Speed up JSONL hydration and snapshot rehydration for portable HotMem mounts.

This PR combines #24 and #25:

• Adds a batch insert path for hydrate so JSONL imports commit once instead of once per memory.
• Adds SQLite-native duplicate skipping through INSERT OR IGNORE and a partial unique content_hash index where possible.
• Preloads existing content hashes so duplicate records are skipped before embedding work.
• Adds hydrate trace counters for parsed rows, loaded rows, skipped duplicates, bytes read, computed embeddings, reused embeddings, and total duration.
• Adds embedding_b64 to snapshots by default so compatible snapshots can hydrate without recomputing embeddings.
• Keeps legacy swap.jsonl files working when embeddings are missing, malformed, or incompatible.
• Preserves ttl_seconds, created_at, metadata, and existing CLI/API behavior.

Why

HotMem's plug-and-play mount story depends on small serialized memory files hydrating quickly into a usable SQLite memory DB. The old path paid avoidable costs: per-row commits, per-row dedupe queries, and repeated embedding computation for known snapshots.

Closes #24.
Closes #25.

Checklist

• Tests pass (uv run pytest)
• Lint passes (uv run ruff check src/ tests/)
• No new external dependencies added to core