Prompt Perturbation Consistency Learning for Robust Language Models