Dynamic Sparse No Training: Training-Free Fine-tuning for Sparse LLMs