GIRAFFE: Representing Scenes as Compositional Generative Neural Feature Fields