Interpreting Neural Policies with Disentangled Tree Representations