Accelerating Scalable Graph Neural Network Inference with Node-Adaptive Propagation