Making Pre-trained Language Models both Task-solvers and Self-calibrators