feat: use shared informers to reduce memory and goroutines

[zachsmith1]

At 200 VCP scale, the dominant cost was informer/listener fanout (O(clusters * resources)) and duplicated controller wiring. Namespace and admission webhook paths were still paying per-cluster overhead, and CRD routing/controller setup had become harder to reason about.

This PR reduces that fanout by moving to shared informer surfaces with cluster-scoped projection, while preserving strict per-cluster isolation semantics.

How It Works

1. Shared Informer + Cluster-Scoped Views (Namespace and Webhook)
   Namespace and webhook managers now rely on shared informer factories and scoped listers/informers instead of spinning per-cluster factories for hot resources.
   scopedFactory implementations project a per-cluster view from shared indexes using cluster-label filtering.
   Result: one shared watch pipeline for key resources, cluster isolation enforced at projection/lister boundaries.

2. List-Backed Webhook Configuration Sources
   Mutating/validating webhook config sources now read via listers and cache accessors by name+resourceVersion.
   They intentionally avoid per-plugin informer handler registration, collapsing listener fanout.
   Result: lower goroutine count and less memory pressure in webhook-heavy paths without changing webhook behavior.

3. CRD Multicluster Routing/Controller Cleanup
   CRD routing is refactored to use unified multicluster handling (serveClusterCRD) and clearer controller wiring.
   Shared CRD projection/runtime logic is used for cluster-aware lookup and status flow, reducing duplicated control paths.
   Result: simpler routing semantics, fewer replicated loops, and better correctness under multicluster churn.
   Impact and Verification

Improves scalability by replacing per-cluster informer/controller duplication with shared pipelines plus cluster-scoped projection.

Preserves isolation: CRUD/list/watch behavior remains per-cluster from caller perspective.
Validated with smoke coverage across RBAC, namespace lifecycle, webhook scoping, CRD establish/discovery/update, and 200 VCP memory runs.

Net effect: materially lower goroutine overhead and improved memory profile consistency at high VCP counts.