On the Tractability of Neural Causal Inference