Learning Choice Functions with Gaussian Processes