Beyond Prompting: Making Pre-trained Language Models Better Zero-shot Learners by Clustering Representations