On Embeddings for Numerical Features in Tabular Deep Learning