chore: upgrade db-retry to 0.8.1 and SQLAlchemy to 2.0.55

[lesnik512]

Raises the db-retry floor from 0.8.0 to 0.8.1 and upgrades SQLAlchemy from 2.0.54 to 2.0.55 in the lock. The sqlalchemy[asyncio]<2.1 pin stays.

In db-retry 0.8.1, Transaction.__aenter__ rolls back the session's transaction on entry when that transaction can no longer continue: SQLAlchemy marked it inactive after a failed flush, or a connection it already holds was invalidated (modern-python/db-retry#75). With 0.8.0, a connection lost on a query or flush after the block made the retried attempt fail with PendingRollbackError. With 0.8.1 the retry starts on a fresh connection and recovers. A live transaction is still joined, so uncommitted work is kept.

The rule that every query goes inside the Transaction block still applies. A query that runs before the block fails the retried attempt before __aenter__ gets a chance to roll back.

Release notes: https://github.com/modern-python/db-retry/releases/tag/0.8.1

SQLAlchemy 2.0.55 (released 2026-10-16) fixes ticket 13673. If a task was cancelled a second time while the pool was closing an asyncio connection invalidated by the first cancellation, the closed connection stayed in the pool, and the next checkout got it and failed on first use. With pool_pre_ping off, only postgres_retry would have caught that. The release also fixes related cancellation handling in the pool (tickets 12710, 13622 and 13678).

Checked locally: just lint, just test and just test-migrations in Docker, with SQLAlchemy 2.0.55 and db-retry 0.8.1 installed in the container.