Towards Physically Interpretable World Models: Meaningful Weakly Supervised Representations for Visual Trajectory Prediction