STAR: Improving Low-Resource Information Extraction by Structure-to-Text Data Generation with Large Language Models