fix: 处理 streaming 响应中 choices 为空的情况

[KangweiZhu]

使用NVIDIA NIM提供的minimax-m2.5 API，运行python llm_client.py 会报

content = chunk.choices[0].delta.content or ''
              ~~~~~~~~~~~~~^^^ IndexError: list index out of range

Debug后发现，一些模型的最后一个chunk会是token usage报告， 其choices为空

ChatCompletionChunk(id='chatcmpl-81e6962c93ce26f7', choices=[Choice(delta=ChoiceDelta(content='性能。祝 你编码顺利！', function_call=None, refusal=None, role=None, tool_calls=None), finish_reason='stop', index=0, logprobs=None, stop_reason=None, token_ids=None)], created=1772561517, model='minimaxai/minimax-m2.5', object='chat.completion.chunk', service_tier=None, system_fingerprint=None, usage=None)

ChatCompletionChunk(id='chatcmpl-81e6962c93ce26f7', choices=[], created=1772561517, model='minimaxai/minimax-m2.5', object='chat.completion.chunk', service_tier=None, system_fingerprint=None, usage=CompletionUsage(completion_tokens=1368, prompt_tokens=43, total_tokens=1411, completion_tokens_details=None, prompt_tokens_details=None))

模型完成响应

<think>The user asks in Chinese: "写一个快速排序算法" i.e., "Write a quicksort algorithm." This is a straightforward request for code. The policy: It's allowed content. The user asks for an algorithm; we can provide code in some language, maybe Python, perhaps also discuss algorithm.

加一条额外判断就可以解决问题。