fix(llm): keep evaluator-only config out of the request

[seancoding-day]

Evaluator config and provider request parameters share one bag — EvaluatorLLMArgs is extra="allow", so both land in model_extra — and everything in it was forwarded to chat.completions.create. The SDK takes no unknown keyword arguments, so configuring a documented knob such as threshold (RAG, agent_eval, instruction_quality) or strictness made every call by that evaluator raise TypeError, which the caller then shapes into an ordinary "evaluation failed" — indistinguishable from a timeout.

What may be sent is now decided by the SDK's own create() signature rather than a list someone has to remember to update. A denylist was tried first and missed two keys on the first pass; the signature travels with the SDK. Keys that are neither request parameters nor registered local knobs are dropped with a warning, so a misspelled key no longer fails silently.

llm_custom_metric forwarded model_extra directly and so bypassed the filter entirely; it now goes through the same accessor. request_timeout and max_retries become configurable per evaluator, the latter reaching the client rather than the request body. Paths that had no timeout before only get one when configured — inventing a default there would fail long calls that work today.