A stream can now declare max_concurrency in config/torque.php: a worker
skips polling that stream while it already has that many of its jobs in
flight, so bulk work (API backfills, scrapes) cannot monopolize the fiber
fleet or hammer an external service, while other streams keep full
throughput. The cap is per worker process (fleet-wide = workers x cap)
and approximate under bursts, overshooting by at most the number of
concurrently polling fibers.