Do 3D Large Language Models Really Understand 3D Spatial Relationships?