Characterizing Vision-Language-Action Models across XPUs: Constraints and Acceleration for On-Robot Deployment