GLaD: Geometric Latent Distillation for Vision-Language-Action Models