Enhancing Multi-hop Reasoning through Knowledge Erasure in Large Language Model Editing