Exploring the Limit of Outcome Reward for Learning Mathematical Reasoning