feat: generic chat_template_kwargs (model config + per-request metadata)

[localai-bot]

What

Closes #10329.

Adds a generic way to pass arbitrary jinja chat-template variables (e.g. Qwen3's preserve_thinking) to backends, so a new template lever no longer needs a hardcoded block in grpc-server.cpp (as enable_thinking in #8973 and reasoning_effort in #10184 each did).

Two sources, no new API surface:

• Model YAML - a new chat_template_kwargs: map (typed):

  name: qwen3
  chat_template_kwargs:
    preserve_thinking: true

• Per-request - the existing (previously unwired) OpenAI metadata field. String values; "true"/"false" are coerced to booleans, anything else stays a string:

  {
    "model": "qwen3",
    "messages": [{"role": "user", "content": "hi"}],
    "metadata": { "preserve_thinking": "true", "enable_thinking": "false" }
  }

How

Precedence (low -> high): config map < server reasoning levers (enable_thinking/reasoning_effort) < per-request metadata.

• core/config - ModelConfig.ChatTemplateKwargs (YAML) + RequestMetadata (request-scoped carrier) + ResolveChatTemplateKwargs(meta) which layers the config map under the coerced metadata and skips the reserved chat_template_kwargs key.
• core/backend/options.go (gRPCPredictOpts) - merges client RequestMetadata over the server-derived levers (so a per-request enable_thinking/reasoning_effort override reaches every backend via the standalone metadata keys), then serialises the resolved map into a single metadata["chat_template_kwargs"] JSON blob, written last so a client cannot clobber it.
• core/http/middleware/request.go - stamps the request metadata onto the per-request config.
• backend/cpp/llama-cpp/grpc-server.cpp - replaces the two per-key enable_thinking/reasoning_effort blocks (streaming + non-streaming) with one generic block that parses the blob and merges every key into body_json["chat_template_kwargs"]. New template levers now need no C++ change.
• Docs + tests.

The standalone metadata["enable_thinking"]/["reasoning_effort"] keys are still emitted (sglang, mlx-vlm, mlx-distributed, vllm-omni read them); other backends receive the new chat_template_kwargs metadata key and harmlessly ignore it. enable_thinking reaches llama.cpp as a real JSON bool (preserving the old == "true" behaviour); reasoning_effort stays a string.

Notes / by design

• A model-YAML chat_template_kwargs value is folded only into the llama.cpp blob, while per-request metadata keys also become standalone gRPC metadata keys (so they reach the Python backends). This asymmetry is intentional: the feature is llama.cpp/jinja-centric, and typed (non-boolean) values are YAML-only.
• This moves enable_thinking/reasoning_effort chat_template_kwargs construction from C++ into Go; llama.cpp output is unchanged, but the gRPC metadata now carries a chat_template_kwargs blob whenever a reasoning lever or kwarg is active.
• The C++ change was reviewed by reading; it is verified by the CI llama.cpp backend build.

Test plan

• go test ./core/config/ ./core/backend/ ./core/http/middleware/ - green (resolver precedence/coercion/reserved-key, gRPCPredictOpts blob + client-override + anti-clobber + omit, middleware metadata wiring).
• golangci-lint (new-from-merge-base) - 0 issues; gofmt clean.
• llama.cpp backend build (CI) - confirms the generic C++ merge compiles.

Assisted-by: Claude:claude-opus-4-8