Pruning Pre-trained Language Models Without Fine-Tuning