HiMTok: Learning Hierarchical Mask Tokens for Image Segmentation with Large Multimodal Model