feat(vllm-cpp): add GLiNER2.5 NER via TokenClassify

[mudler-agent]

feat(vllm-cpp): add GLiNER2.5 NER and kev-compatible SystemOne API

Wires the vllm-cpp backend to the C ABI NER surface (vllm_gliner_ner,
ABI v27) so LocalAI can serve zero-shot named entity recognition, then
adds the full kev-compatible SystemOne structured-extraction API on top.

Depends on mudler/vllm.cpp#3224, which adds the GLiNER2.5 model, the
DeBERTa v2 encoder, the boundary head, the C ABI vllm_gliner_ner
function, and the /v1/ner and /v1/systemone server endpoints.

NER backend

• backend/go/vllm-cpp/backend.go: TokenClassify method on *VllmCpp calls
  vllm_gliner_ner with text and labels, copies the C-owned entity array
  into protobuf TokenClassifyEntity messages, and frees the result. Uses
  in.Labels when non-empty, falling back to configured ner_labels then
  defaultNerLabels.
• backend/go/vllm-cpp/govllmcpp.go: cNerEntity and cNerResult Go POD
  mirrors matching the C structs; vllmGlinerNer and vllmNerResultFree
  purego bindings; abiVersion bumped 26 -> 27.
• backend/go/vllm-cpp/options.go: ner_labels, ner_threshold, ner_max_width
  parsed from engine_args.
• backend/go/vllm-cpp/vllmcpp_test.go: ABI version assertion updated
  26 -> 27; cNerEntity and cNerResult struct mirror tests added.
• backend/backend.proto: repeated string labels = 4 added to
  TokenClassifyRequest for per-request zero-shot label selection.
• pkg/grpc/interface.go: ClassifyModel interface (opt-in extension to
  AIModel, follows the Embedding pattern).
• pkg/grpc/server.go: TokenClassify gRPC handler.
• core/config/backend_capabilities.go: vllm-cpp backend declares
  MethodTokenClassify and UsecaseTokenClassify.
• core/backend/token_classify.go: Labels field in TokenClassifyOptions,
  TokenClassifyWithLabels method on TokenClassifier, ModelTokenClassify
  passes Labels in the gRPC request.

SystemOne API (kev-compatible)

Mirrors the API from https://github.com/jaredpalmer/kev. Three POST
endpoints run zero-shot NER over rendered state text and build
kev-compatible answers for three question types: noul (binary entity
presence), choice (pick one option), and score (pick one level).

• core/schema/systemone.go: SystemOneRequest, SystemOneQuestion,
  SystemOneResponse, SystemOneAnswer, SystemOneEntity, SystemOneUsage,
  SystemOnePermuteRequest, SystemOnePermuteRun, SystemOnePermuteResponse.
• core/http/endpoints/localai/systemone.go: SystemOneEndpoint,
  SystemOnePermuteEndpoint (re-runs one choice question under n_perm
  option orders with seeded RNG), SystemOneSeparateEndpoint (N
  independent NER passes, one per question). Helpers: renderState,
  softmax, choiceConfidence, scoreConfidence, r2, parseSystemOneRequest,
  buildSystemOneAnswer, resolveClassifier — ported from kev/api.py.
• core/http/routes/systemone.go: RegisterSystemOneRoutes wiring the
  three POST routes.
• core/http/app.go: route registration alongside RegisterPIIRoutes.
• docs/content/features/vllm-cpp.md: NER section + SystemOne API section
  documenting the endpoints, question types, and engine_args keys.

The helper functions are mirrored in vllm.cpp's api_server.cpp so both
servers produce the same answer shape.

Following-Agents-Protocol: true
AI-Assisted: true
Assisted-by: AGENT:regolo/glm5.2 [maki]