Forgetting before Learning: Utilizing Parametric Arithmetic for Knowledge Updating in Large Language Models