GLISTER: Generalization based Data Subset Selection for Efficient and Robust Learning