Interpretable Neural Causal Models with TRAM-DAGs