fix(distributed): exclude pinned models from cluster eviction and idle scale-down

[Dennisadira]

Description

Addresses the cluster-side eviction gap in #11101.

pinned: true was only honoured by the per-node WatchDog (idle eviction, LRU limit, memory reclaimer, group eviction). Every distributed eviction path was pinned-blind:

• SmartRouter.EvictLRU → FindLRUModel
• SmartRouter.evictLRUAndFreeNode (locked SQL selection)
• ReplicaReconciler.scaleDownIdle

Since eviction is gated on in_flight = 0, a pinned model became eviction-eligible the instant each response completed. Under capacity pressure that surfaces exactly as the report describes: the backend is freed immediately after every request, with the reporter's settings (watchdog off, max_active_backends: 0, reclaimer off) creating no watchdog at all, so the local exclusion never had a chance to apply.

Fix

Wire the model config loader into the router and reconciler through a new PinnedModelResolver seam (mirrors ConcurrencyConflictResolver):

• EvictLRU passes the pinned set into FindLRUModel's query, so the next-oldest unpinned model is selected instead of the attempt failing
• evictLRUAndFreeNode filters pinned models inside its locked selection
• scaleDownIdle skips pinned models entirely

Deliberate teardown (admin unload, model delete, node drain) intentionally still applies to pinned models, as does dead-row reaping.

Tests

• New core/services/nodes/pinned_eviction_test.go regression specs, verified to fail with the exclusion disabled
• GetPinnedModelNames specs in the config loader tests
• Upstream revision_eligibility_test.go updated for the new FindLRUModel signature

Docs

docs/content/advanced/vram-management.md documents that pinned now also protects models in distributed mode.

Notes for reviewers

The pkg/grpc build on master currently needs make protogen-go first (committed pb bindings lag backend.proto); unrelated to this change.

Fixes #11101

[localai-org-maint-bot]

This does not compile: the FindLRUModel signature change misses a caller.

core/services/nodes/interfaces.go and registry.go move to FindLRUModel(ctx, nodeID, excludeModels []string), and the in-package callers and fakes are all updated, but tests/e2e/distributed/node_registration_test.go:219 still calls the two-argument form:

vet: tests/e2e/distributed/node_registration_test.go:219:66: not enough arguments in call to registry.FindLRUModel
	have (context.Context, string)
	want (context.Context, string, []string)

Reproduced on the PR head with go vet ./tests/e2e/distributed/.... There is no build tag on that file, so it is a plain build break of the package; it slipped through because the only check that ran on this PR was DCO.

Fix is to pass nil at that call site, matching how reconciler_inflight_leak_test.go and revision_eligibility_test.go were already updated in this PR.

The rest of the change looks good to me: the PinnedModelResolver seam, the reconciler exemption, and the docs update are all consistent, and the comment about deliberate teardown deliberately bypassing the exemption is the right call.