Mixtures of Laplace Approximations for Improved Post-Hoc Uncertainty in Deep Learning