On the Stability and Scalability of Node Perturbation Learning