Input-Output Alignment for Efficient 3D Manipulation Learning with Vision-Language Models