ThinkingConfig lacks budget_tokens

ThinkingConfig (the thinking argument of chat.completions.create) only declares type. Its docstring says it is for Gemini models.

When a chat request is converted for Claude, the gateway also reads thinking.budget_tokens. Callers who pass {"type": "enabled", "budget_tokens": 2048} get a type error, although the field works.

Fix: add budget_tokens: int to ThinkingConfig, and make the docstring say which models read it.