When Does Bounded-Optimal Metareasoning Favor Few Cognitive Systems?