Hypothesis Transfer Learning with Surrogate Classification Losses: Generalization Bounds through Algorithmic Stability