Towards an Information Theoretic Framework of Context-Based Offline Meta-Reinforcement Learning