fix(utils): align Qwen tokenization for parallel tool calls

[looput]

The fix groups consecutive tool messages before passing them to apply_chat_template, preserving the formatting defined by the native Qwen2 chat template.

The original issue was caused by calling apply_chat_template separately for every message in gen_multi_turn_loss_mask_qwen, without preserving the context between consecutive tool messages.

For a parallel tool-calling sequence:

assistant(tool_calls=[A, B])
tool(result=A)
tool(result=B)

The Qwen2 chat template is designed to group consecutive tool responses into a single user block:

<|im_start|>user
<tool_response>
result=A
</tool_response>
<tool_response>
result=B
</tool_response><|im_end|>

However, the original implementation rendered each tool message independently. Since each message appeared to be both the first and last tool message, it produced:

<|im_start|>user
<tool_response>
result=A
</tool_response><|im_end|>
<|im_start|>user
<tool_response>
result=B
</tool_response><|im_end|>

This caused:

• token_ids to differ from the native Qwen chat-template output;
• redundant <|im_start|>user and <|im_end|> tokens to be inserted;
• the training input format to differ from the model’s actual inference format;
• token positions and loss-mask alignment to potentially become incorrect.

The issue only affects parallel tool calling with consecutive role="tool" responses. A single assistant message containing multiple tool_calls, or sequential assistant → tool → assistant → tool interactions, does not trigger this problem.