Anisotropic, Sparse and Interpretable Physics-Informed Neural Networks for PDEs