cluster recovery: fix resource limits and node1 memory

- nvidia quota: requests.memory 8Gi → 12Gi (unblock cuda-validator)
- calibre: startup probe initial_delay 60→120s, timeout 1→5s,
  wait_for_rollout=false (DOCKER_MODS install takes 10+ min)
- immich ML: memory 2Gi → 4Gi (OOMKilled loading CLIP models)

Also done outside TF (not in this commit):
- node1 VM: 16 GiB → 24 GiB RAM (Proxmox)
- tigera-operator: kubectl patch 128→256Mi
- nvidia-driver-daemonset: kubectl patch 1→4Gi memory
- kyverno reports-controller: kubectl patch 128→256Mi
- CNPG operator: kubectl rollout restart

stacks/immich/main.tf

@@ -513,10 +513,10 @@ resource "kubernetes_deployment" "immich-machine-learning" {
           resources {
             requests = {
               cpu    = "100m"
-              memory = "2Gi"
+              memory = "4Gi"
             }
             limits = {
-              memory = "2Gi"
+              memory = "4Gi"
               "nvidia.com/gpu" = "1"
             }
           }