On Structured Prediction Theory with Calibrated Convex Surrogate Losses