Cross-Hand Latent Representation for Vision-Language-Action Models