Model selection by penalization in mixture of experts models with a non-asymptotic approach