feat: add JobExhaustedError and reclaim event for orphan tracing

Add a dedicated JobExhaustedError class (src/errors.ts) for claim-time
exhaustion: raised when a job is claimed past its maxAttempts without ever
completing — typically a worker killed mid-run (deploy/OOM/eviction) that
was reclaimed by the stale lease. Carries queue, jobId, attempts,
maxAttempts, and lastError (prior attempt's recorded failure, or null if
orphaned before any error), so the dead/jobError events are self-explanatory
without digging into the database. Exported from the package public API.

Add a reclaim event to PrismaQueueEvents, emitted by requeueStale whenever
it resets stale claims. Reports the id and stuckForMs of each reclaimed job,
connecting the lease recovery to the later dead event on the same id.

Rewrite requeueStale() to use a single atomic
  WITH stale AS (SELECT … FOR UPDATE SKIP LOCKED)
  UPDATE … SET processedAt = NULL FROM stale RETURNING id, prevProcessedAt
statement instead of updateMany — this avoids a snapshot/update race,
returns exactly the rows that were reset (not an ambiguous count), and uses
SKIP LOCKED to side-step any concurrent dequeue rather than blocking on it.