refactor: use single poller with semaphore-based capacity control

Previously, capacity=N spawned N independent polling goroutines, each
making FetchTask RPCs to the Gitea server concurrently. This caused
unnecessary connection load on the server proportional to the runner's
capacity setting.

Replace the N-goroutine model with a single polling loop that uses a
buffered channel as a semaphore to control concurrent task execution.
The poller acquires a capacity slot before fetching; when at capacity,
it blocks without issuing RPCs. Fetched tasks are dispatched to
independent goroutines that release their slot on completion.

Also fix a pre-existing bug in Shutdown() where the timeout branch
used a blocking receive on p.done instead of a non-blocking select,
which prevented shutdownJobs() from ever being called on timeout.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

internal/pkg/metrics/metrics.go

@@ -89,7 +89,7 @@ var (
 		Namespace: Namespace,
 		Subsystem: "poll",
 		Name:      "backoff_seconds",
-		Help:      "Last observed polling backoff interval. With Capacity > 1, reflects whichever worker wrote last.",
+		Help:      "Last observed polling backoff interval in seconds.",
 	})
 
 	JobsTotal = prometheus.NewCounterVec(prometheus.CounterOpts{