fix(memory): serialize concurrent add() calls with asyncio.Lock

[octo-patch]

Fixes #1381

Problem

When ReActAgent runs with parallel_tool_calls=True, multiple _acting coroutines invoke AsyncSQLAlchemyMemory.add() concurrently. Without serialization:

1. Coroutine A checks: "Does message X exist?" → No (not yet committed)
2. Coroutine B checks: "Does message X exist?" → No (same DB state)
3. Both decide to insert message X
4. One succeeds; the other fails with:

sqlalchemy.exc.IntegrityError: (pymysql.err.IntegrityError) (1062,
"Duplicate entry '...' for key 'PRIMARY'")

The same race condition affects _get_next_index(), where two concurrent calls can read the same max index, leading both to attempt inserts with the same index value.

Solution

Add an asyncio.Lock to AsyncSQLAlchemyMemory and wrap the entire check-then-write section of add() (duplicate check → index read → message insert → mark insert → commit) in async with self._lock:. This serializes concurrent add() calls per memory instance, eliminating the TOCTOU race condition.

Testing

• Root cause confirmed: asyncio.Lock prevents concurrent coroutines from interleaving their read-check-insert-commit sequences.
• Only per-instance serialization is needed: callers that use separate AsyncSQLAlchemyMemory instances are not affected.
• The lock does not affect the common single-coroutine case (no contention).