Concept-modulated model-based offline reinforcement learning for rapid generalization