Computational Limitations in Robust Classification and Win-Win Results