Unleashing the Power of Large Language Models in Zero-shot Relation Extraction via Self-Prompting