On the Opportunities and Challenges of Offline Reinforcement Learning for Recommender Systems