Model Predictive Control with Differentiable World Models for Offline Reinforcement Learning