Small Language Models are Good Too: An Empirical Study of Zero-Shot Classification