Sparse and Faithful Explanations Without Sparse Models