Towards Accurate and Calibrated Classification: Regularizing Cross-Entropy From A Generative Perspective