Add screenshot tool, integrate with vision

[ErikBjare]

This could allow for running/testing GUI applications, and more E2E multimodal behavior.

Xvfb in itself might be a bad idea (although good for running in headless/CI), could just take screenshots directly (also more platform independent).

Not sure how to add input, but keyboard-focused input should be possible.

Vision tracking issue: #50

[ErikBjare]

Finished the browser version of this, but yet to automatically include the resulting screenshot into the conversation: #52