Matryoshka Query Transformer for Large Vision-Language Models