Wake idle workers when jobs are queued

[rksharma-owg]

Closes #517

Summary

• add a PostgreSQL notification channel for queued jobs
• wait for either a job notification or the existing polling backstop when the worker is idle
• emit notifications transactionally for normal, deduplicated, extraction, RSS, and manual requeue paths
• add a database-backed regression test for the notification path

Validation

• cargo fmt --all --check
• UTOPIA_DATABASE_URL=postgres://utopia:utopia@localhost:1517/utopia UTOPIA_TEST_REQUIRE_DB=1 cargo test --workspace
• cargo clippy --workspace --all-targets -- -D warnings
• ./node_modules/.bin/vite build

The repository's pnpm build wrapper could not complete because this environment's pnpm policy requires explicit approval for the esbuild install script; the direct Vite build passed.

[WaylandYang]

Merged, with two changes pushed to your branch first so it could land today:

• The idle wait sleeps when the listener fails. On a recv error the loop went straight back to claim_one; a database that accepts claims while the listener cannot reconnect (an exhausted pool is one way) would have turned that into a tight loop of one claim and one warning per iteration. The wait is now one function, jobs::wait_for_work, and an error sleeps the poll interval before the next claim.
• The test exercises what the worker does. The original asserted that a notification arrives. The new file drives wait_for_work itself, which is what the loop calls, and pins three rules: an enqueue ends the wait with Notified inside the poll interval; a row inserted in an open transaction wakes nothing until commit, which is the ordering hazard A queued job waits two seconds for a worker that is doing nothing #517 names; with nothing queued the poll arm still fires. It stays on the wait function rather than run_worker because a live worker on a shared test database would claim other tests' jobs.

The notification sites are exactly as you wrote them, all five. One comment added: the listener keeps a pool connection for good, so the pool has one fewer to hand out. Thanks for the complete coverage; that is what decided between the two PRs for #517.