fix(sweeper): Phase 3H — recover from BLPOP-to-claim race

[smallchungus]

Closes a real correctness gap.

Before this, a worker that crashed between BLPOP (job ID gone from Redis) and ClaimJob (PG row not yet flipped to running) left the job stuck — status=queued with no last_error, invisible to the existing sweeper paths (reviveOrphans wants status=running; promoteDelayed wants last_error set).

• store.ListStaleQueuedJobs: SELECT WHERE status=queued AND last_error IS NULL AND created_at < now() - threshold
• store.PoolForTest exposes the pool for tests that need raw SQL (backdating created_at to simulate stale rows)
• sweeper.SweepOnce now also calls repushStale, which LPUSHes any stale-queued job IDs back onto their stage queue. Duplicate pushes are safe — ClaimJob is atomic; a worker that picks up an already-claimed job ID gets ErrJobNotClaimable and moves on
• cmd/sweeper reads STALE_QUEUED_THRESHOLD_SEC env var (default 60)
• .env.example documents the new var

3 new tests (1 store + 2 sweeper). All existing tests still green. Loadtest still passes (1643 jobs/s).