Multi-Stage Manipulation with Demonstration-Augmented Reward, Policy, and World Model Learning