fix: use dtype instead of deprecated torch_dtype

[xyf5432]

Summary

config.torch_dtype and the torch_dtype keyword argument were deprecated in transformers 4.56 (PR #39782) and replaced by dtype. This PR updates the two from_pretrained calls in train.py (ShowUIForConditionalGeneration line 331 and Qwen2VLForConditionalGeneration line 344) to select the keyword from the installed transformers version using packaging.version, so transformers < 4.56 keeps working. All other arguments (_attn_implementation, quantization_config, device_map, lm_skip_layer, lm_skip_ratio, ...) are unchanged.

Changes

• train.py: add a _dtype_kwargs helper that returns {"dtype": ...} on transformers >= 4.56 and {"torch_dtype": ...} otherwise; both from_pretrained calls pass it as **_dtype_kwargs(torch_dtype).

Test

• The modified file compiles (python -m py_compile).
• On transformers >= 4.56 the calls use dtype and emit no deprecation warning; on older versions they pass torch_dtype unchanged.

Fixes #99