Uncertainty-Aware Step-wise Verification with Generative Reward Models