A Close Look into the Calibration of Pre-trained Language Models