On Learning Representations for Tabular Data Distillation