Edge-Cloud Routing for Text-to-Image Model with Token-Level Multi-Metric Prediction