Differential replay of reward and punishment paths predicts approach and avoidance