Debiasing Watermarks for Large Language Models via Maximal Coupling