Fixes several issues/bugs with VisionAgent:

• Fixes bug where if you pass a list of size > 2 to edit_vision_code it will crash because you've passed more than one user role message in a row
• Fixes bug where it would hallucinate when using view_media_artifact
• Fixes issue where it would try to call view_media_artifact on the users behalf
• Fixes issue where it could not write numpy arrays with write_media_artifact
• Adds rule for ensuring it will only call edit_vision_code on code generated from generate_vision_code, as opposed to using edit_code_artifact
• Adds rule for ensuring it will not continue to modify code for things the user did not ask for. For example, sometimes after generating vision code it would modify it to include visualizations
• Added more test cases for functions that modify code input and fixed some edge cases