On Uncertainty, Tempering, and Data Augmentation in Bayesian Classification