Optimize model serving with GKE Inference Gateway