Leashing the Inner Demons: Self-Detoxification for Language Models