Support NVIDIADriver reconciliation without a ClusterPolicy

[karthikvetrivel]

Stacked on #2571 (kv-gpuclusterconfig-crd). This PR's diff is just the four commits below; GitHub will retarget it onto main once #2571 lands.

Summary

Builds on the GPUClusterConfig controller so the NVIDIADriver operand can reconcile on its own, without a ClusterPolicy:

• Resolves the driver host root from the info catalog rather than requiring the ClusterPolicy CR.
• The controller reconciles driver state when no ClusterPolicy is present.
• GPU node labeling driven by the GPUClusterConfig controller.
  • TODO: Incorporate Create new controller for all node labeling operations #2569 (move node labeling into the dedicated node-labeling controller).
• A disabled GPUClusterConfig re-reconciles so it picks up later ClusterPolicy removal.

Testing

Validated live on a single-node Tesla T4 cluster (Kubernetes 1.34, containerd 2.2), plus unit coverage for host-root sourcing.

• I created a GPUClusterConfig on a cluster with no ClusterPolicy present and verified the node got the relevant nvidia.com/* labels, the NVIDIADriver-owned driver DaemonSet came up, and the DRA kubelet-plugin published its DeviceClasses and ResourceSlice. I then ran a pod with a DRA claim and verified nvidia-smi worked against the operator-installed driver.
• With a ClusterPolicy present, I verified the NVIDIADriver stays disabled. I deleted the ClusterPolicy and verified the NVIDIADriver recovered to ready on its own within about a minute.
• With neither a ClusterPolicy nor a GPUClusterConfig in the cluster, I verified the NVIDIADriver goes notReady with "no ClusterPolicy or GPUClusterConfig object found in the cluster". I created a GPUClusterConfig and verified the NVIDIADriver recovered.
• I confirmed a GPUClusterConfig that's disabled by an existing ClusterPolicy keeps re-reconciling while disabled, so it notices and takes over once that ClusterPolicy is later removed.