Add PostgreSQL adapter as alternative storage backend

[josharsh]

Summary

Engram currently uses SQLite for everything (storage, FTS5, sqlite-vec). This works great for single-process and embedded use cases, but teams running multiple agent instances need a shared database.

A PostgreSQL adapter would enable multi-process and distributed deployments.

Design considerations

• The MemoryStore class in src/store/store.ts is the abstraction boundary. A PG adapter should implement the same interface.
• PostgreSQL has tsvector for full-text search (replaces FTS5)
• PostgreSQL has pgvector extension for vector similarity search (replaces sqlite-vec)
• Migrations should be separate from SQLite migrations
• The adapter should be an optional dependency (don't break the zero-dependency SQLite default)

Suggested approach

1. Extract a StorageAdapter interface from the current MemoryStore methods
2. Keep existing SQLite implementation as the default
3. Add a new pg-store.ts that implements the same interface using pg + pgvector
4. Let users choose via config: { adapter: 'sqlite' | 'postgres', connectionString: '...' }

This is a significant contribution

This is a larger piece of work suited for someone comfortable with both SQLite and PostgreSQL. Happy to pair on the interface design before implementation starts — open a draft PR or comment here.

Acceptance criteria

• Same test suite passes against both SQLite and PostgreSQL adapters
• PostgreSQL adapter uses pgvector for similarity search
• PostgreSQL adapter uses tsvector for full-text search
• SQLite remains the default with zero extra dependencies
• Documentation for PostgreSQL setup