Analyzing Cost-Sensitive Surrogate Losses via $\mathcal{H}$-calibration