Self-prompted Chain-of-Thought on Large Language Models for Open-domain Multi-hop Reasoning