Variational approximations of empirical Bayes posteriors in high-dimensional linear models