fix(longhorn): isolate GPU worker from general storage scheduling

- Remove expand-replicas-job (blindly forced all volumes to 3 replicas,
  ignoring StorageClass settings)
- Add diskSelector: 'storage' to longhorn and longhorn-cnpg StorageClasses
  so replicas only land on CP nodes (cp-1, cp-2, cp-3)
- Tag all CP node disks with 'storage' via PostSync job
  (disk names are runtime-discovered, can't hardcode in Node CRs)
- Disable scheduling on worker-1 Node CR — only longhorn-llm-local
  (diskSelector: 'llm') can use it
- worker-1 is GPU-only: llm-models and comfyui use dedicated SCs

k8s/infra/longhorn/kustomization.yaml

@@ -8,11 +8,11 @@ resources:
   - longhorn-servicemonitor.yaml
   - longhorn-taint-toleration.yaml
   - longhorn-nodes.yaml
-  - expand-replicas-job.yaml
+  - longhorn-tag-disks-job.yaml
   - patch-csi-tolerations-job.yaml
   - longhorn-add-disks-job.yaml  # Add extra disks to talos-cp-2
 # Longhorn deployed via bootstrap script or Helm.
 # These manifests configure it: unified StorageClass (default, 3 replicas),
 # Prometheus ServiceMonitor, taint toleration for control-plane nodes, explicit
 # Node CRDs for cp-2/cp-3, CSI plugin tolerations, and a PostSync hook Job
-# that ensures all existing volumes have 3 replicas.
+# that configures storage for the cluster.