Causal Lifting of Neural Representations: Zero-Shot Generalization for Causal Inferences