fix(langchain): stop dropping temperature on gpt-5.x models [PC-4988]

[tudormatei1]

langchain-openai strips temperature for every gpt-5* model unless reasoning effort is explicitly the string "none". Unset effort is None, not "none", so an agent's configured temperature never reached the dotted variants (gpt-5.2, gpt-5.4), which default to effort none and do accept it. Nothing errored or warned, so the value just vanished between agent.json and the request.

Confirmed the models accept it: a Responses call to gpt-5.4 with temperature: 0.6 and no reasoning_effort returns 200 and echoes back temperature: 0.6 with reasoning.effort: "none".

Affects the langchain package only. Gpt5TemperatureMixin overrides both sites langchain strips at (validate_temperature and _construct_responses_api_payload) on UiPathChatOpenAI and UiPathAzureChatOpenAI, and defers to langchain everywhere it is already right, so base gpt-5, pro variants, models with reasoning explicitly enabled, and models flagged shouldSkipTemperature by discovery keep losing the parameter as before.

Start reading at gpt5_keeps_temperature. The mixin has to go before the vendor class on the MRO, and the _get_request_payload override reinstates a value the layer below deliberately removed, which reads oddly but is unavoidable: the second strip is a module-level function with no hook between it and the request.

Two canary tests assert the upstream bug is still present on unpatched ChatOpenAI. They fail when the upstream fix ships and name the files to delete, so this cannot rot silently.

Verified end to end through the agents runtime against gpt-5.4. ruff check, ruff format --check, pyright and the openai/sampling tests are clean; the pre-existing failures elsewhere in pytest tests need gateway credentials and are unchanged by this branch.

Upstream: langchain-ai/langchain#35423, fix pending in langchain-ai/langchain#35424.