Understanding Finite-State Representations of Recurrent Policy Networks