fix(neo4j): cascade delete vectors from Qdrant when deleting memory nodes

[AYin-Z]

Description

delete_node_by_prams() only ran DETACH DELETE on Neo4j, leaving orphan vectors behind in Qdrant. Every memory deletion via the API created a leaked vector in Qdrant with no corresponding Neo4j node.

Root Cause

The write path (add_node / add_nodes_batch) is symmetric — it writes to both Neo4j and Qdrant in the same method. But the delete path only touched Neo4j:

# Before: count then delete Neo4j only
count_query = f"MATCH (n:Memory) WHERE ... RETURN count(n) AS node_count"
...
session.run("MATCH (n:Memory) WHERE ... DETACH DELETE n", **params)
# ← No vec_db.delete() call

Note that clear() and the newer delete_node_by_mem_cube_id() both correctly call vec_db.delete() — this was simply an omission in delete_node_by_prams.

Fix

1. Replace count_query with an ids_query that collects node IDs
2. Delete from Neo4j (same as before)
3. Call self.vec_db.delete(deleted_ids) after successful Neo4j deletion

Verification

• BaseVecDB.delete(ids) is an abstract method with working implementations in both QdrantVecDB and MilvusVecDB
• self.vec_db is initialized in the constructor and used elsewhere in the same class
• The fix gracefully falls back with a warning log if Qdrant cleanup fails

[Memtensor-AI]

Automated Test Results: PASSED

Cloud test-engine rerun against dev-v2.0.22 completed successfully.

• Run: tr-c67f3a71-38d on cloud test-engine 10011
• memos_python_core/changed-python-source: 2 passed, 0 failed, 0 skipped

Manual code review is still required before merge.