Backward Imitation and Forward Reinforcement Learning via Bi-directional Model Rollouts