Neural Generators of Sparse Local Linear Models for Achieving both Accuracy and Interpretability