fix(ctl): honour temperature/max_tokens declared under spec.models

[jordanauge]

Problem

resolve_model_name() already accepts the model name from either spec.llm or spec.models[0], but engine_factory read temperature and max_tokens from spec.llm only.

Any manifest using the spec.models[] shape therefore silently fell back to the defaults (max_tokens=2000, temperature=0.7). A manifest declaring:

spec:
  models:
    - id: main
      model: vertex_ai/gemini-2.5-flash
      temperature: 0.0
      max_tokens: 8000

was actually run at max_tokens=2000 / temperature=0.7, with no warning.

Impact

This is easy to miss on short answers, but it is fatal on reasoning models, which bill internal reasoning against the same completion budget. Observed on a multi-round negotiation with Gemini 2.5 Flash:

• completion_tokens: 1996, of which reasoning_tokens: 1916 and text_tokens: 80 → the answer is truncated mid-sentence
• in the worst case completion_tokens: 0 with finish_reason: "stop" → an empty completion, because reasoning consumed the whole (unexpectedly small) budget

Raising max_tokens in the manifest had no effect, which made this quite hard to diagnose: the declared value never reached the engine.

Change

• _resolve_sampling_param(): symmetric spec.llm → spec.models[0] lookup for temperature and max_tokens, mirroring what resolve_model_name already does for the model name. spec.llm still wins when both are set, and the existing defaults are unchanged when neither is.
• _resolve_model_option() + LiveLlmEngine.reasoning_effort: optional reasoning_effort pass-through, so a manifest can bound internal reasoning on models that bill it against max_tokens. Defaults to None, so the request payload is unchanged unless a manifest opts in.

Compatibility

No behaviour change for manifests using spec.llm. Manifests using spec.models[] start getting the values they already declared.