Imagination Policy: Using Generative Point Cloud Models for Learning Manipulation Policies