Robust scalable initialization for Bayesian variational inference with multi-modal Laplace approximations