On the Calibration of Pre-trained Language Models using Mixup Guided by Area Under the Margin and Saliency