Co-training Improves Prompt-based Learning for Large Language Models