Add optional Milvus vector store backend

[zc277584121]

Summary

• add an optional MilvusVectorStoreBackend implemented with MilvusClient
• support dense search, native BM25 full-text search, structured metadata filters, schema reuse validation, and lifecycle operations
• preserve the ascending semantic-distance and descending keyword-score contracts across Milvus Lite, Milvus Server, and Zilliz Cloud
• keep LanceDB as the default and expose Milvus through the milvus installation extra and documented URI/token settings

Usage

pip install -e ".[milvus]"

store = VectorStore(
    backend_factory=lambda dimension: MilvusVectorStoreBackend(
        collection_name="memory_entries",
        vector_dimension=dimension,
        uri="./milvus.db",
    )
)

The same backend accepts a Milvus Server or Zilliz Cloud endpoint through MILVUS_URI and MILVUS_TOKEN.

Validation

• python -m pytest -q tests/test_vector_store_backend.py tests/test_milvus_vector_store_backend.py (16 passed)
• focused Ruff lint and format checks
• git diff --check
• Milvus Lite provider coverage for all retrieval paths, score ordering, safe filters, schema reuse, dimension mismatch, lifecycle operations, and hybrid retrieval
• Zilliz Cloud smoke test with a unique temporary collection and cleanup

No breaking changes; the default LanceDB path is unchanged.