feat(vm): add GPU DRA resource claim support

[danilrwx]

Description

Add support for attaching physical GPU devices to virtual machines via Kubernetes DRA (Dynamic Resource Allocation).

A new spec.gpuDevices field lets a user request a GPU by product model. The virtualization controller generates a DRA ResourceClaimTemplate per device with a CEL selector matching the requested productName, a physical device type, and the absence of a sharing strategy. The kvbuilder renders the corresponding DRA resource claims and GPU devices into the KubeVirt VirtualMachine.

Key pieces:

• VirtualMachine.spec.gpuDevices[] (name, model), MaxItems: 16, +listType=map keyed by name.
• GPU feature gate (alpha, locked off in CE).
• GPUResourceClaimHandler creates/updates/deletes owned ResourceClaimTemplates and cleans up orphans; a ResourceClaimTemplate watcher enqueues the owning VM.
• GPUDevicesValidator rejects GPU devices unless the GPU feature gate is enabled and the gpu.deckhouse.io DeviceClass exists.
• vmchange comparator marks GPU changes as requiring restart (AwaitingRestartToApplyConfiguration).
• Generated names: claim/GPU name = gpu-<name>, device request name = gpu-<name>, ResourceClaimTemplate = <vm>-<name>.

Depends on deckhouse/3p-kubevirt#130 for KubeVirt to recognize the Deckhouse GPU DRA attributes.

Why do we need it, and what problem does it solve?

Users running ML/rendering workloads need to attach a physical GPU to a VM. Today there is no way to do this through the VirtualMachine API.

A model-based request keeps VM manifests portable: the user asks for a GPU class (e.g. NVIDIA H100) and lets DRA + the scheduler pick a concrete, exclusive, passthrough-capable device on a suitable node — instead of pinning the VM to a specific node, PCI address, or GPU UUID.

What is the expected result?

1. Enable the GPU feature gate in ModuleConfig virtualization and ensure a GPU DRA provider and the gpu.deckhouse.io DeviceClass are installed.
2. Create a VM with spec.gpuDevices:

   spec:
     gpuDevices:
       - name: gpu0
         model: NVIDIA H100

3. The controller creates a ResourceClaimTemplate and the VM schedules on a node with a matching exclusive physical GPU.
4. Changing spec.gpuDevices on a running VM sets AwaitingRestartToApplyConfiguration and applies only after a restart.

Checklist

• The code is covered by unit tests.
• e2e tests passed.
• Documentation updated according to the changes.
• Changes were tested in the Kubernetes cluster manually.

Changelog entries

section: core
type: feature
summary: "Attach physical GPU devices to virtual machines via DRA by requesting a GPU product model in spec.gpuDevices."