Testing the Generalization of Neural Language Models for COVID-19 Misinformation Detection