On Fundamental Limits of Robust Learning