DPO Meets PPO: Reinforced Token Optimization for RLHF