feat(node-agent): Node Agent CDI Simulator

[roma-glushko]

What This PR Does

Ported the CDI logic as a new node agent Simulator.
Cleaned up nvml-mock scripts from ported logic.

Why

A part of MEP0003.

Checklist

• Commits are signed off (git commit -s)
• Tests pass (go test -v -race ./...)
• Linter passes (make lint-fix)
• New code has SPDX license headers
• Documentation updated (if applicable)
• CHANGELOG.md updated (if user-facing change)

---

_{Stack created with GitHub Stacks CLI • Give Feedback 💬}

[giuliocalzo]

Non-blocking suggestion: consider giving the CDI simulator a logger and emitting one line per Apply, the way I ended up doing it for rdmaplugin in #734.

The reasoning is the same for both. agent.reconcile logs only failures, so a simulator that succeeds is invisible, and FileSource emits an update on subscribe and then only when the config hash changes — so Apply runs once at startup and then rarely. One Info per Apply is therefore bounded, and it's the only evidence the surface was actually published.

That matters more for CDI than for most surfaces: when the spec is missing or stale, the symptom shows up far away, as containerd refusing container creation or injecting nothing, and the first question is whether the agent wrote the spec at all and with how many devices.

Something like:

func New(log *slog.Logger) *Simulator { return &Simulator{log: log} }

func (s *Simulator) Apply(_ context.Context, h *host.Host, state *agent.State) error {
    nvidiaPath := filepath.Join(h.Run, nvidiaSpecFile)
    nriPath := filepath.Join(h.Run, nriSpecFile)
    if err := writeSpec(h, nvidiaPath, buildNvidiaSpec(state)); err != nil {
        return fmt.Errorf("nvidia.yaml: %w", err)
    }
    if err := writeSpec(h, nriPath, buildNRISpec(state)); err != nil {
        return fmt.Errorf("nvml-mock-nri.yaml: %w", err)
    }
    s.log.Info("wrote CDI specs", "simulator", name,
        "nvidia", nvidiaPath, "nri", nriPath, "devices", len(state.Devices))
    return nil
}

rdmaplugin also carries a patched field to distinguish a real write from a no-op, but that only exists because it reads the node before patching. host.WriteFile rewrites unconditionally, so there's no equivalent here and I wouldn't add change detection just for the log line.

Happy to fold this into #734 instead if you'd rather keep this PR as a straight port.

The base branch was changed.

[ArangoGutierrez]

/lgtm