Learning Calibrated Uncertainties for Domain Shift: A Distributionally Robust Learning Approach