Fact-and-Reflection (FaR) Improves Confidence Calibration of Large Language Models