Extreme Low-Bit Clustering for Large Language Models via Knowledge Distillation