Causal Abstractions of Neural Networks