Improving Diversity of Commonsense Generation by Large Language Models via In-Context Learning