Probabilistically Robust Learning: Balancing Average- and Worst-case Performance