Learning beyond Teacher: Generalized On-Policy Distillation with Reward Extrapolation