Multi-Task Fusion via Reinforcement Learning for Long-Term User Satisfaction in Recommender Systems