On the (im)possibility of branch-and-bound search-to-decision reductions for approximate optimization.