fix(memory): release SQLite engine config entries when engines are collected

[adityasingh2400]

SQLAlchemySession._configure_sqlite_engine remembers that an engine already has the SQLite connect listener by storing id(engine.sync_engine) in a class-level set, but that set holds no reference to the engine, so the entry outlives the object. CPython reuses addresses, so a later engine allocated at the same address takes the early return and never gets PRAGMA busy_timeout or PRAGMA journal_mode = WAL, which is exactly the configuration _run_sqlite_write_with_retry exists to complement. The same set also grows for the lifetime of the process, which from_url makes easy to hit because it creates and discards an engine per session.

This registers a weakref.finalize callback that discards the entry when the sync engine is collected, matching the guard MongoDBSession already documents for its own id()-keyed registry. Two regression tests cover it: one asserts the entry is gone after the engine is collected, and one asserts the registry returns to its baseline size after 25 short-lived from_url sessions. Both fail on main and pass with the fix, and uv run pytest tests/extensions/memory/test_sqlalchemy_session.py is green at 36 passed along with make format, make lint, and make typecheck.