fix(llm): stop sending DashScope-only enable_thinking to official OpenAI

[Voyagerroc-Lab]

Fixes #38

What was wrong

thinking=False paths injected extra_body["enable_thinking"] = False unconditionally. enable_thinking is a DashScope/Qwen extension — the official OpenAI API rejects it with 400 Bad Request: Unknown parameter: 'enable_thinking', so every storyboard/dialogue extraction run against api.openai.com failed (as reproduced in the issue, where the reporter also verified that switching to reasoning_effort="none" unblocks the official endpoint).

The same unconditional bind also existed in AgentBase.__init__, so every agent sent enable_thinking to the official API regardless of its value.

What changed

Three call sites now branch on the target endpoint:

• backend/app/services/llm/resolver.py (_build_chat_openai_model)
• backend/app/services/llm/runtime.py (build_default_text_llm_sync)
  • official OpenAI (api.openai.com, or no custom base_url) → kwargs.setdefault("reasoning_effort", "none")
  • DashScope-style compatible endpoints → unchanged extra_body["enable_thinking"] = False
• backend/app/chains/agents/base.py (AgentBase.__init__) — the unconditional self._model.bind(extra_body={"enable_thinking": ...}) is replaced by _thinking_extra_body(model, self.enable_thinking), which returns None (nothing bound) for official-OpenAI targets and the previous body for compatible endpoints, using the model's openai_api_base.

setdefault keeps any user-configured reasoning_effort from model.params intact.

How did you verify?

• python -m py_compile passes for all three touched files.
• Official-OpenAI case (base_url empty or api.openai.com): no enable_thinking key reaches the wire; reasoning_effort="none" is sent only when the caller did not set one.
• DashScope case (custom base_url): behavior is byte-identical to before.

中文摘要：enable_thinking 是 DashScope 扩展参数，官方 OpenAI API 会返回 400。现在三个调用点都会按端点区分：官方 OpenAI 使用 reasoning_effort="none"，DashScope 兼容端点保持原有 enable_thinking 行为。