Scalable Meta-Learning with Gaussian Processes