fix(controller): drop model label from Deployment selector to make modelRef mutable (closes #301)

[Defilan]

Closes #301.

Bug

Deployment.Spec.Selector.MatchLabels included inference.llmkube.dev/model: <modelRef-value>. Kubernetes treats Deployment selectors as immutable after creation, so editing spec.modelRef on a running InferenceService:

1. Accepted the CR change (kubectl returned success)
2. Every subsequent reconcile failed at the apiserver with `field is immutable`
3. The Pod kept running the old model with no error surfaced to the user (the failures only showed up in controller logs)

This blocked common workflows like swapping a quant level or rolling to a different upstream fork.

Fix

Split the label set:

• deploymentSelectorLabels (new helper): app + inference.llmkube.dev/service. Both are derived from isvc.Name and never change over the InferenceService's lifetime, so the selector is byte-stable across reconciles and the apiserver `Update` succeeds.
• Full label map (including the model label) still ships on Deployment.ObjectMeta.Labels and Pod template labels, so `kubectl get pods -l inference.llmkube.dev/model=` keeps working as a filter.

When the user now edits modelRef, the selector stays identical, the controller updates the Pod template (new model label, new init container args, new model path), and Kubernetes does a standard rolling update.

Test plan

• Two regression tests in `inferenceservice_deployment_test.go`:
  • selector must NOT contain `inference.llmkube.dev/model`
  • swapping `modelRef` on the same InferenceService produces Deployments whose selectors are deeply equal (this IS the apiserver immutability constraint, asserted at the source)
• `go test ./internal/controller/... -count=1` passes
• `golangci-lint run ./internal/controller/...` clean (0 issues)
• No CRD changes; no chart changes

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!