Fix CUDA kernel error on Turing GPUs when resetting LoRA (#418)

[lvyonghuan]

• Detect Turing GPUs and use safe LoRA reset method
• Replace reset_lora() with update_lora_params({}) for Turing
• Add CUDA synchronization and fallback handling
• Maintain backward compatibility for non-Turing GPUs

This fixes the 'invalid configuration argument' error at misc_kernels.cu:281 that occurred when removing or switching LoRA models on RTX 2080 Ti and other Turing architecture GPUs.

Solution

• Detect Turing GPUs using is_turing() utility
• Use model.update_lora_params({}) instead of model.reset_lora() for Turing GPUs
• Add CUDA synchronization for stability
• Implement multi-level fallback handling
• Maintain backward compatibility for non-Turing GPUs

Testing on RTX 2080 Ti 22g.

Fixes #418