On the Representational Capacity of Recurrent Neural Language Models