Fix Gemma SFT masking and DPO preference formatting

[houyongsheng]

Summary

• keep DPO prompts and completions separate using the conversational preference format expected by TRL
• load the saved SFT processor when available and set DPO beta explicitly
• mask every assistant span in multi-turn SFT records while enforcing the configured maximum length
• use FastVisionModel and UnslothVisionDataCollator for the optimized multimodal path
• add dependency-free helpers and regression tests for formatting, truncation, and masking

Testing

• Python unittest suite: 7 passed
• AST parsing: all 6 gemma-trainer asset scripts passed
• committed diff whitespace check passed

No model weights were downloaded and no GPU training run was performed.