dsh-vision-assist: Adding Multimodal Recognition Capabilities to Non-Vision Models

The `dsh-vision-assist` plugin intercepts the LLM invocation boundary to address the issue in DeepSeek Harness where pure-text models cannot handle image inputs. When the routed model does not support images, the plugin converts the image into text instructions and triggers the selected multimodal model to perform recognition. It supports hot configuration updates, allowing you to switch models or adjust timeout settings without restarting. The installation command is `dsh plugin --profile web add github:wjling/dsh-vision-assist`, and it requires DSH version >= 0.1.7-alpha.1.