Scalable Training of Inference Networks for Gaussian-Process Models