conformance: add driver-runtime-management check (CNCF MUST at v1.38) #3607

Part of #3601.

Summary

Add a driver-runtime-management conformance check. driver_runtime_management is SHOULD at v1.37 and graduates to MUST at v1.38 (kubernetes-sigs/ai-conformance KAR-0001).

Upstream contract (TestAcceleratorDriverRuntimeManagement)

1. The platform advertises the driver version for the node: ResourceSlice device attribute driverVersion preferred, GFD label nvidia.com/cuda.driver-version.full as fallback. No advertised version is a warning, not a failure.
2. A pod built from a vanilla image (ubuntu:22.04 by default) that requests one accelerator sees exactly the requested devices, the runtime configuration injected the vendor tools (nvidia-smi present), the driver initializes and reports a version, and that version matches the advertised one at shared precision.

Today

The deployment phase proves the GPU Operator is healthy (operator-health, gpu-operator-version, check-nvidia-smi using the operator's own images). Nothing exercises the vanilla-image injection path or compares the running driver with what the platform advertises.

Change

• New check in validators/conformance/, catalog entry, added to every recipe that lists the conformance MUST set.
• Image configurable for air-gapped registries, matching the upstream flag.
• Evidence: advertised version source and value, pod output, comparison.

Acceptance

• Passes on the current EKS, GKE, AKS, OKE and NVKE recipes; fails when the toolkit injection is disabled.