Global inducing point variational posteriors for Bayesian neural networks and deep Gaussian processes