Toward Breaking Watermarks in Distortion-free Large Language Models