smolagents adds vision capabilities with VLM support

Vision capabilities have been added to HuggingFace's smolagents library, enabling VLM (Vision Language Models) to be used natively in agent pipelines.

There are two main ways implemented to pass images:

• Initial image passing: Multiple images can be passed at once via the images argument when calling the agent.run method. This is useful for tasks that analyze visual elements within long PDF documents.
• Dynamic step-by-step passing: Images can be dynamically added at each step of the agent using step_callbacks. This allows a web browsing agent to check the changed screen in real time after taking an action and proceed with the task.

With this update, agents can now directly recognize a web page's layout, icons, and color information, which are difficult to grasp through text extraction alone, enabling more sophisticated autonomous web browsing.