Post-hoc Uncertainty Learning using a Dirichlet Meta-Model