Detoxifying Large Language Models via Knowledge Editing