Real-World Offline Reinforcement Learning from Vision Language Model Feedback