Reinforcement Learning with Token-level Feedback for Controllable Text Generation