Poisson empirical Bayes estimation: When does g-modeling beat f-modeling in theory (and in practice)?