fix: correct four review findings in dequeue and job handling

- Include queue in scheduled-job unique key (@@unique([queue, key, runAt]))
  and update upsert to use queue_key_runAt so separate queues sharing the
  same key+runAt no longer collide.
- Gate deleteOn: failure|always deletion on isFinished so jobs scheduled
  for retry are not deleted before their next attempt in both transactional
  and non-transactional dequeue paths.
- Replace undefined-sentinel success/error detection with a DequeueOutcome
  discriminated union so workers returning void correctly emit the success
  event.
- Type PrismaJob.fetch() as Promise<DatabaseJob | null> to match the actual
  findUnique return when the row has been deleted.