Fix: Remove --apply-chat-template from Qwen3-235B SFT script

[kaysonyu]

Summary

This PR removes the --apply-chat-template flag from scripts/run-qwen3-235B-A22B-sft.sh to align it with the fix made in PR #1307.

Root Cause

The --apply-chat-template flag should not be used in SFT training because:

1. MultiTurnLossMaskGenerator.get_loss_mask() expects messages in list format (list of dicts with role and content keys)
2. When --apply-chat-template is enabled during the Dataset loading phase, it converts messages to a string format via tokenizer.apply_chat_template()
3. This causes get_loss_mask() to fail since it needs to iterate through messages to generate proper loss masks for assistant responses
4. The tokenizer.apply_chat_template() should be called internally by MultiTurnLossMaskGenerator, not during data loading

Changes

• Comment out --apply-chat-template in scripts/run-qwen3-235B-A22B-sft.sh (line 52)