Inference Latency Prediction at the Edge