Riemannian Laplace approximations for Bayesian neural networks