Expand DRA Driver Integration for GPU Operator

[karthikvetrivel]

Purpose

Umbrella branch for NVIDIA DRA driver integration into GPU Operator, building on Chris's foundation work in #1541. Refreshes embedded DRA assets to the newer driver layout and adds the operator-side glue needed to make DRA usable.

Change Log

dd7cf4d:

• Refresh embedded DRA manifests to the new driver layout: split kubelet-plugin RBAC/SAs into separate nvidia-dra-driver-kubeletplugin and compute-domain-daemon-service-account identities, update controller + kubelet-plugin workloads, drop the legacy entrypoint ConfigMap, and switch to the dra-driver-nvidia-gpu image.
• CRD packaging: refresh ComputeDomain, add new ComputeDomainClique, wire both into the Helm CRD upgrade hook, Dockerfile, and OLM bundle.

3be3986:

• Adds three rules to gpu-operator's ClusterRole: pods/patch (consumed by compute-domain-daemon), resource.nvidia.com/computedomaincliques CRUD (the new sub-resource), and coordination.k8s.io/leases (DRA controller leader election).

f9c67dd:

• New draDriver.featureGates field (map[string]bool) on ClusterPolicy. Operator renders it as a FEATURE_GATES env var on all three DRA driver containers (controller, gpus kubelet plugin, compute-domains kubelet plugin) — same channel the upstream k8s-dra-driver-gpu Helm chart uses, identical Key=Value, format with trailing comma.
• Keys are sorted alphabetically before rendering (upstream chart doesn't — Go map iteration is randomized per call). This avoids spurious pod restarts caused by the env-var string shuffling on each reconcile pass. The binary's parser is order-insensitive, so observable behavior is identical to upstream.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.