fix: stop per-node pod watchers when nodes are deleted

[gnuletik]

The nodes informer only handled Add/Update, so when a node was deleted its per-node pod watcher was never torn down. Each leaked node kept a blocked goroutine, a running pod informer (with its own clientset and watch connection), and a podStore entry holding every cached pod. In an autoscaled cluster with churning nodes this grows unbounded (~10 MiB/day observed in prod).

Track each node's stop channel and add a DeleteFunc that closes it via a single stopNodeWatcher helper (close owned solely there, guarded by map membership under isHandling, so exactly-once).

Also fixed two related issues in the same path:

• Reserve the podStore slot synchronously under the lock so a burst of node updates can't spawn two watchers for one node.
• Log-and-clean instead of logrus.Fatalf on pod-informer creation failure; a Fatal in the goroutine took down the whole controller and left every node tainted.