LoRS: Efficient Low-Rank Adaptation for Sparse Large Language Model