Toward a Theory of Causation for Interpreting Neural Code Models