Globally optimal score-based learning of directed acyclic graphs in high-dimensions