Submodular Data Selection and Augmentation for Resource-Efficient Learning