[Langchain-Openai] ChatOpenAI Forwards Unsupported Sampling Parameters To Gpt-6-Astra Via Responses API

The Responses API request construction in langchain_openai does not validate or filter sampling parameters based on model capabilities. For gpt-6-astra (a reasoning model), temperature, top_p, and top_logprobs are not supported by the OpenAI API, but langchain-openai forwards them unchanged, leading to server-side rejection.

1. Set OPENAI_API_KEY.\n2. Instantiate ChatOpenAI with model='gpt-6-astra', use_responses_api=True, max_tokens=32, max_retries=0.\n3. Call invoke() with temperature=0.2, top_p=0.5, or top_logprobs=1.\n4. Observe HTTP 400 errors: 'Unsupported parameter: temperature', 'Unsupported parameter: top_p', 'logprobs are not supported with reasoning models.'

The proposed fix relies on a hardcoded whitelist of unsupported models, which requires manual updates when new models with similar restrictions are released. It only validates when use_responses_api=True; the Chat Completions API may have different support, so the mapping must be maintained per API type. Additionally, if a model name is passed as an alias or via a custom resolver, the mapping lookup could miss it. To rollback, remove the _validate_sampling_params call and the mapping; the original behavior (passing parameters through) will be restored, but users will again encounter server-side 400 errors for these models.