Adversarial Robustness of Prompt-based Few-Shot Learning for Natural Language Understanding