Improve Visual Threat Model Evaluator

[heyitsaamir]

1. Old Visual Threat Model Evaluator was using multiple agents (questioner, answerer, answer evaluator). The new one is a bit different. It uses very specific specs along with instructions on how to evaluate it. It also suggests a very clear outline on how to receive responses (using tool-calling).
2. Update autogen to latest version. This includes a change that I introduced to autogen (Support async nested chats microsoft/autogen#3309). This allows nested chats to not take over the whole process and instead runs them async on the server.
3. Instead of using a global flag, I made it into a state (with useXML the default). If you send /useVisual it swaps the mode and /useXml swaps it back