Linear-Core Surrogates: Smooth Loss Functions with Linear Rates for Classification and Structured Prediction