feat(memory): add async user metadata extraction pipeline

- Add MetadataExtractor to collect user-related statements post-dedup
  and extract profile/behavioral metadata via independent LLM call
- Add Celery task (extract_user_metadata) routed to memory_tasks queue
- Add metadata models (UserMetadata, UserMetadataProfile, etc.)
- Add metadata utility functions (clean, validate, merge with _op support)
- Add Jinja2 prompt template for metadata extraction (zh/en)
- Fix Lucene query parameter naming: rename `q` to `query` across all
  Cypher queries, graph_search functions, and callers
- Escape `/` in Lucene queries to prevent TokenMgrError
- Add `speaker` field to ChunkNode and persist it in Neo4j
- Remove unused imports (argparse, os, UUID) in search.py
- Fix unnecessary db context nesting in interest distribution task

api/app/celery_app.py

@@ -111,6 +111,9 @@ celery_app.conf.update(
         # Clustering tasks → memory_tasks queue (使用相同的 worker，避免 macOS fork 问题)
         'app.tasks.run_incremental_clustering': {'queue': 'memory_tasks'},
 
+        # Metadata extraction → memory_tasks queue
+        'app.tasks.extract_user_metadata': {'queue': 'memory_tasks'},
+
         # Document tasks → document_tasks queue (prefork worker)
         'app.core.rag.tasks.parse_document': {'queue': 'document_tasks'},
         'app.core.rag.tasks.build_graphrag_for_kb': {'queue': 'document_tasks'},