Do We Still Need Complex Vision-Language Pipelines? Researchers from ByteDance and WHU Introduce Pixel-SAIL—A Single Transformer Model for Pixel-Level Understanding That Outperforms 7B MLLMs