feat(rust): steering, capture, and named steering modules in the rust frontend

[RhizoNymph]

What

Adds per-request activation steering and end-to-end activation capture to the Rust frontend, plus named steering modules (startup load + runtime management), across both inbound surfaces (OpenAI HTTP /v1/completions + /v1/chat/completions, and the served gRPC Generate service).

Per-request steering

• New request fields: steering_vectors / prefill_steering_vectors / decode_steering_vectors (packed base64 wire format matching the Python OpenAI API), steering_name.
• Packed format (dtype/shape/layer_indices/data/scales) decoded into the inline SteeringVectorSpec engine-core resolves (float32/float16/bfloat16/float64, per-row scales). Layer keys serialize as integer msgpack keys (Python dict[int, ...]). steering_name lowers to steering_module_ref = (name, 1.0).

Named steering modules

• --steering-modules name=path loads module JSON at startup and broadcasts to the engine workers (register_steering_modules + pre_materialize_steering_module).
• Runtime endpoints re-broadcast on change: GET / POST / DELETE /v1/steering/modules. Per-request steering_name validated up front (unknown → invalid_request / NotFound); mutations serialized by a lock.

Activation capture (end-to-end)

• Request side: capture spec accepted on HTTP + gRPC, forwarded verbatim into SamplingParams.capture; engine-core's offline admission resolves prefix-cache flags.
• Output side (the key fix): Python's EngineCoreOutput is an array_like tuple with capture_results at index 7; the Rust EngineCoreOutput predated that field, so capture-enabled outputs misaligned/failed to decode. Added a CaptureResult type and inserted capture_results at the correct tuple index, threaded it (alongside kv_transfer_params) through llm → text → chat, and surfaced it on the non-streaming completion/chat responses and the gRPC FinishInfo — coercing each consumer payload to an object (mirror of Python's _capture_result_to_response_payload). Streaming responses omit capture_results, matching Python (capture still executes; consumers write out of band).
• The python_compat.py fixture mirror was updated to include capture_results, keeping the Rust↔Python wire-format test faithful; a new wire test pins the index-7 position.

Why

The Rust frontend replaces vLLM's Python OpenAI entrypoint; these are the steering/capture surfaces clients rely on. Per-request resolution/admission stays in the engine-core; the frontend decodes the wire format, forwards, manages the named-module registry, and surfaces capture results.

Notes / deferred

• Capture-spec validation parity: the Python entrypoint's synchronous _admit_capture spec validation is deferred to the engine-core here, so a malformed capture spec surfaces as an engine error rather than a 400 up front.

See rust/docs/features/steering_capture.md for the full data flow (incl. the capture return path), file map, and invariants.