dsh-image-mmx: Adding image understanding to DSH text models

`dsh-image-mmx` is a community plugin for DeepSeek Harness (MIT licensed, maintained by fengs2021) designed to address the limitation of pure text models in handling image inputs. Its core mechanism is to automatically invoke MiniMax VLM (mmx) for visual recognition when users send images in a session, and inject the recognition results into the model context, enabling text models like DeepSeek to understand image content. Main features include: supporting the sending of messages containing images to avoid errors; automatically saving images to the session workspace and executing `mmx vision describe`; replacing image blocks on the model side with path text and detailed descriptions (preserving visual elements, code, error messages, etc.), while the user interface continues to display original image thumbnails that can be enlarged; and supporting the parallel and independent recognition of multiple images.