Reverse engineering recurrent neural networks with Jacobian switching linear dynamical systems