Optimizing Long-term Value for Auction-Based Recommender Systems via On-Policy Reinforcement Learning