SAM3D-Guided Object-Centric Representation Alignment for Vision-Language-Action Models