LaMem-VLA: Dual Latent Memory in Vision-Language-Action Models for Robotic Manipulation | ResearchPod