Cross-Table Pretraining towards a Universal Function Space for Heterogeneous Tabular Data