Pruning Pre-trained Language Models with Principled Importance and Self-regularization