On the convergence problem in mean field games: a two state model without uniqueness