Generative Verifiers: Reward Modeling as Next-Token Prediction