A Tunable Loss Function for Robust Classification: Calibration, Landscape, and Generalization