Poincaré-Bendixson Limit Sets in Multi-Agent Learning