ViVa: A Video-Generative Value Model for Robot Reinforcement Learning