Question about implementation of ThinkingModelTipsScheduler

Checklist / 检查清单

• I have searched existing issues, and this is a new bug report. / 我已经搜索过现有的 issues，确认这是一个新的 bug report。

Bug Description / Bug 描述

I have a few questions regarding the ThinkingModelTipsScheduler implementation:

1. In run(), for round_output, we only keep the last thinking block, yet current_request.messages still contains thinking blocks from all previous rounds during inference. Should we make current_request.messages only retains the latest thinking as well?
   https://github.com/modelscope/ms-swift/blob/d9010ed937680bfb23520ab1e9c2ad85f826be4a/swift/rollout/multi_turn.py#L471-L472

2. In _build_messages(), when no template is used, the appended element seems incorrect. Based on the format of messages, should it be: {'role': 'assistant', 'content': assistant_no_think} ?
   https://github.com/modelscope/ms-swift/blob/d9010ed937680bfb23520ab1e9c2ad85f826be4a/swift/rollout/multi_turn.py#L572-L575

Thanks for your help!

How to Reproduce / 如何复现

running the framework's script: examples/train/grpo/external/vllm_multi_turn.sh and check related variables.

Additional Information / 补充信息

No response