On the Efficient Minimization of Classification Calibrated Surrogates