Efficient Robotic Policy Learning via Latent Space Backward Planning