Text this: Zero-Shot Prediction of Conversational Derailment With Large Language Models