feat: add the OCI image source to ModelArtifact

[thxCode]

What type of PR is this?

/kind enhancement
/kind api-change
/area worker
/area testing

What this PR does / why we need it:

Adds an image source to ModelArtifact: weights identified by a digest-pinned OCI image
reference (registry/repository@sha256:<64 hex>) and delivered by mounting a Kubernetes image
volume — kubelet pulls the pinned image on the node that needs it, and nothing enters the
model-manager plugin's cache chain.

• Admission accepts the member behind a version capability gate: the union names image,
  the reference must pin a digest (a mutable tag would let one artifact deliver different weights
  on different pulls), patterns are refused because an image is mounted whole, and creation is
  refused on an apiserver below the ImageVolume default-on floor (1.35), where the volume field
  would be dropped silently. The gate reads the startup version snapshot through a swappable
  function; an unreadable version refuses. ModelPrefetch refuses an image artifact: there is
  nothing to warm.
• The artifact reconciler resolves an image source with no network: Resolved=True on the
  first pass with a claim-shaped status — no revision, no manifest digest (the pinned reference is
  the whole identity; the KV identity falls back to the artifact UID), no status.nodes, no
  revalidation.
• Both consumers render the delivery: a ModelDeployment and an Instance mount one image
  volume, read-only, at /var/lib/gpustack/model — no cache volume, no download environment, no
  --revision, no ephemeral-storage raise — and status.model.delivery reports Image. An
  Instance pinned to a node checks that node's kubelet (≥ 1.35) and containerd (≥ 2.1) before
  rendering. The soft placement preference names the nodes whose Node.status.images list the
  reference, hostname-sorted and capped, never a filter.
• Docs: a new page owns the source (digest contract, the build recipe measured byte-equal to
  its hub commit, version floors, double storage 1.79×/2.00×/2.26× and why the CRI size must not
  be used for capacity, the three image-GC rules, registry mirrors); model-artifact.md, the docs
  index and the skill routing tables follow. No new Setting.
• e2e: case-112 proves the cold-mount path on a live cluster — the tag refusal, the
  claim-shaped resolution, and an Instance reading the image's fixture byte for byte — with the
  fixture generated and pushed into an in-cluster registry over the node's loopback, so no
  containerd configuration is touched.

Deviation from the adjudicated draft, recorded in the spec: AC13's preference ordering is
hostname-sorted, name-only — Node.status.images carries no referenced-now or last-used signal,
so the digest preference's serving-now ordering has nothing to sort on. Weight, cap and softness
stand.

Which issue(s) this PR links to:

None

Special notes for your reviewer:

• The chart matrix was run locally against all seven node images (v1.23.17 → v1.35.5, each
  installing gpustack/gpustack-operator:chart-test built from this branch; v1.33.12's first two
  attempts failed on Docker Hub anonymous rate limiting — the chart's mirrored dependency images
  pulled from docker.io — and passed with the images preloaded into the nodes, zero docker.io
  pull events in the final run).
• make generate was verified in a clean checkout: zero diff.

Does this PR introduce a user-facing change?

ModelArtifact gains an `image` source: weights delivered as a digest-pinned OCI image reference
through a Kubernetes image volume, on Kubernetes 1.35+ with containerd 2.1+. Admission refuses
tag references and clusters below the version floor; an Instance pinned to a node that cannot
run image volumes waits and names the node.

[gpustack-code-review]

🔍 OpenCodeReview found 3 issue(s) in this PR.

• ✅ Successfully posted inline: 0 comment(s)
• 📋 Routed to summary by policy: 3 comment(s)

---

📄 pkg/worker/webhooks/worker/model_prefetch.go (L112-L116)

⚠️ GitHub could not post this as an inline comment: Routed to summary (severity low · category test)

Reminder per repo rules (not a claim about this PR's contents, since generated and test files are excluded from review): this webhook change must be accompanied by the regenerated zz_generated.webhooks.go output (make generate) and a regression test in the same PR. Likewise, the controller changes in pkg/worker/controllers/worker (image delivery rendering, runtime floor check, placement preference) must ship same-PR regression tests.

---

📄 api/worker/v1alpha1/model_artifact.go (L110-L110)

⚠️ GitHub could not post this as an inline comment: Routed to summary (severity low · category documentation)

Reminder (per project rules, not a claim about presence/absence): this API change (new optional Image source member and the ModelDeploymentModelDeliveryImage constant) is additive and backward compatible — no renames, type changes, or removed markers — but it must ship with regenerated make generate output (zz_generated deepcopy/CRD manifests), which is excluded from this review. Please confirm it was regenerated in this PR.

---

📄 pkg/worker/webhooks/worker/model_artifact.go (L284-L284)

⚠️ GitHub could not post this as an inline comment: Routed to summary (severity low · category test)

Reminder (per project rules, not a claim about presence/absence): the new validation paths here — digest pinning (missing @, malformed digest, empty/whitespace reference), pattern refusal on image sources, and the version-gate refusal — are new bug surface; a regression test covering them should ship in this same PR. Also, the webhook change must be accompanied by the regenerated make generate output (zz_generated.webhooks.go), which is excluded from this review.