CrossJEPA vs. JEPA-WM: Advancing Efficient 3D Learning via Cross-Modal Latent View Prediction. JEPA.