fix: register session listeners in framework extension configs (#709)

[hasansezertasan]

Summary

Fixes #709 — framework extension configs (Litestar, Starlette, Sanic, Flask, async and sync each) override create_session_maker without delegating to super(), silently skipping AsyncFileObjectListener / SyncFileObjectListener, touch_updated_timestamp, and AsyncCacheListener / SyncCacheListener registration. User-visible effect: file uploads via FileObject / StoredObject persist metadata to the database but the actual file content is silently discarded.

• Fix: Each override now delegates to super().create_session_maker() so the middle-layer listener wiring runs. Framework-specific pre-steps (notably Starlette/Sanic/Flask's engine_instance caching) are preserved verbatim — the principle is add listener wiring, preserve every other observable behavior.
• Scope: 8 source files (4 extensions × async + sync). FastAPI inherits from Starlette so gets the fix for free (no override exists there).
• Tests: 24 new unit tests in 5 files mirror the existing mocking pattern in tests/unit/test_config/test_async_config.py, locking the listener-registration contract per extension so future overrides cannot silently drop it again.

Design notes

• Why mock-based unit tests and not integration tests? The base class's integration tests already cover end-to-end listener behavior. The bug is that listeners weren't registered at all in subclasses — a unit-test question. Mocks catch exactly that failure mode loudly and specifically (call_count == 0 vs expected 6), whereas an integration test would fail as a confusing silent missing-file, which is exactly the failure mode Bug: Litestar SQLAlchemyAsyncConfig.create_session_maker() does not register FileObject session listeners #709 reported.
• Why engine_instance pre-step preserved in Starlette/Sanic/Flask? Analysis shows it's technically redundant with get_engine()'s internal memoization, so the overrides could be deleted outright. But "minimal harm" mandates preserving exact call sequences in case downstream users have subclassed or monkey-patched in unexpected ways. The tests pre-populate config.engine_instance = MagicMock() to isolate listener-count assertions from the aiosqlite dialect's internal event.listen calls.
• Atomic commits: 18 small commits tell the full story: 5 failing-test commits (red) → 8 super() delegation commits (green) → 1 test-isolation commit → 3 pre-step restoration commits → 1 test strengthening commit. Each commit is individually revertable.

Test Plan

• 24 new unit tests pass (listener registration contract per extension)
• Full tests/unit/ suite — no regressions (pre-existing failures in test_repository.py and test_utils/test_fixtures.py confirmed unrelated by checking out base commit)
• Reviewer manually verifies MCVE from Bug: Litestar SQLAlchemyAsyncConfig.create_session_maker() does not register FileObject session listeners #709 now produces "file_exists_on_disk": true
• Reviewer confirms the "minimal harm" design approach (preserve engine_instance pre-step) matches project preference

Draft status

Drafting for reviewer discussion on:

1. Whether to keep the conservative engine_instance pre-step or delete the non-Litestar overrides outright (simpler, provably safe per get_engine() memoization analysis).
2. Whether to add CHANGELOG entries or leave them to the release-bump process.

🤖 Generated with Claude Code

[hasansezertasan]

Superseded by #712 — same commits, renamed branch (feat/hasansezertasan/a → fix/extension-create-session-maker-listeners). This PR was auto-closed by GitHub when the source branch was renamed because cross-fork PRs don't preserve rename linkage. Please review #712 instead.