Implement embedded vector search fallback via SQLite BLOB storage

[bug-ops]

Parent: #740 (P0)

Problem

External vector database dependency complicates single-binary and edge deployment. When unavailable, system falls back to keyword-only search losing semantic capability entirely.

Solution

Implement in-process cosine similarity over SQLite BLOB storage as an alternative VectorStore backend:

• Store embeddings as little-endian f32 BLOB arrays in SQLite
• Compute cosine similarity in Rust at query time
• Implement VectorStore trait for the new backend
• Select backend via config: memory.vector_backend = "qdrant" | "sqlite"
• Qdrant remains the production-grade option; SQLite-embedded enables zero-dependency mode

Affected crates

• zeph-memory (new SqliteVectorStore implementing VectorStore trait)

Acceptance criteria

• SqliteVectorStore passes all existing VectorStore trait tests
• Hybrid search works identically with both backends
• Config toggle between backends
• Documentation updated