fix: migrate to async SQLAlchemy and harden runtime robustness

[mbuffa]

• Replace sync SQLAlchemy engine/Session with async engine + AsyncSession (sqlite+aiosqlite) so DB calls no longer block the event loop in async routes and the worker
• Add SQLite PRAGMA busy_timeout=30000 to prevent "database is locked" errors under concurrent writes; also set synchronous=NORMAL and foreign_keys=ON
• Wrap blocking YouTubeTranscriptApi.fetch() in asyncio.to_thread() so transcript fetching no longer stalls every concurrent SSE stream
• Fix SSE endpoint pinning a DB connection for the whole stream lifetime; now uses a short-lived session scoped to the slug lookup only
• Add explicit db.rollback() before retry in worker exception path, preventing PendingRollbackError from masking the real failure
• Replace swallowed exceptions with logger.exception() throughout worker and summarizer; sanitize client-facing error messages (no raw exception text leaked to browser)
• Bound _progress/_listeners memory: prune empty listener lists on unsubscribe; schedule eviction of completed video state after a 60 s grace period via call_later
• Simplify worker pool: remove misleading MAX_CONCURRENT+4/Semaphore pattern — spawn exactly MAX_CONCURRENT workers, one job per worker
• Track worker tasks in a module-level set (prevents premature GC); add stop_workers() called from lifespan for clean shutdown on Ctrl-C
• Update conftest and test_routes to use async fixtures and AsyncSession
• Add aiosqlite>=0.20 and greenlet>=3.0 dependencies