Contrastive Demonstration Tuning for Pre-trained Language Models