From Dense to Sparse: Contrastive Pruning for Better Pre-trained Language Model Compression