fix: make graph add relationship writes transactional

[wayyoungboy]

Summary

Partially addresses #1003.

This PR fixes the independent atomicity part of the graph add() issue by putting relationship deletes and relationship inserts in the same SQLAlchemy transaction.

Changes:

• wraps the add() delete+add relationship write stage in engine.begin()
• passes the active transaction connection into _delete_entities() and _add_entities()
• uses the transaction connection for relationship deletion, relationship existence checks, relationship insertion, and final entity-name reads
• re-raises delete failures inside the transaction path so failures roll back instead of being silently treated as zero deletes
• preserves the existing non-transaction client path for direct helper calls
• adds unit coverage for transaction connection propagation, relationship delete, and relationship insert paths

The check-then-create duplicate entity race should be handled separately after #1002 lands, because the correct unique key depends on graph_entities.user_id and should be (user_id, name).

Validation

• python -m pytest tests/unit/test_oceanbase_graph.py -q
• python -m pytest tests/unit -q
• python -m flake8 src/powermem/storage/oceanbase/oceanbase_graph.py tests/unit/test_oceanbase_graph.py --select=F601,F821,E999
• git diff --check