Multimodal Embodied Agents: A Unified Capability-centric Perspective from Computer-Use to Robot-Use