fix: model dtype is not same as lora dtype in FSDP train

[0hujun]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

As described in #182，in ascend npu, model param dtype is bf16， but create lora param dtype is fp32 as default，that got AssertionError: FSDP expects uniform original parameter dtype but got FSDP expects uniform original parameter dtype but got {torch.bfloat16, torch.float32}
So, when lora param has created, convert all LoRA parameters to the base model dtype.

Experiment results

Train fine as usual

[gemini-code-assist]

Code Review

This pull request introduces the _ensure_lora_dtype method to align LoRA parameter data types with the base model, ensuring compatibility with FSDP2. Review feedback suggests several improvements, including more robust detection of the base data type to handle mixed precision, narrowing exception handling, and wrapping parameter updates in torch.no_grad() for safety.