Knowledge Rumination for Pre-trained Language Models