Flexible Blood Glucose Control: Offline Reinforcement Learning from Human Feedback