Exploring the Potential of Large Language Models as Predictors in Dynamic Text-Attributed Graphs