Why do tree-based models still outperform deep learning on typical tabular data?