On the Generalization of SFT: A Reinforcement Learning Perspective with RewardRectification