projectquota: return errors from StartWatcher instead of klog.Fatalln

[tamalsaha]

Problem

ProjectQuotaReconciler.StartWatcher called klog.Fatalln in three places — a nil reconciler, an unknown resource Kind, and a failed Watch:

if r.ctrl == nil {
    klog.Fatalln("ProjectQuota reconciler is not setup yet!")
}
...
klog.Fatalln(err)

StartWatcher is invoked at runtime from the resource-discovery poller (graph.PollNewResourceTypes). A klog.Fatalln there calls os.Exit, so any one of these conditions on any single discovered resource type takes down the entire kube-ui-server process instead of failing just that watch.

Fix

• StartWatcher now returns an error for each of the three conditions (with the offending resource/GVK included in the message).
• The poller logs the error via klog.ErrorS and continues, consistent with the rest of the resilient discovery loop (cf. Isolate per-connection failures in graph reconciler #422 "Isolate per-connection failures").

No behavior in this path should be able to crash the aggregated apiserver.

Testing

• go build ./... passes.
• go vet and gofmt clean on the changed files.
• grep klog.Fatal over pkg/ returns nothing.