Large Language Models can be Strong Self-Detoxifiers