Model-tuning Via Prompts Makes NLP Models Adversarially Robust