Unexpected behavior when messaging the LLM during streaming

[alexc19113800]

When you interrupt the LLM while it is generating, the extension behaves erratically. The output of the model becomes a part of the user's prompt and the generation spirals out of control. The model starts to continuously generate large amount of text and usually stops after many generations

From a glance it looks like the problem might be multiple streams being triggered and logging to the same conversation simultaneously.

May want to cancel the request when a new prompt comes in?

To recreate:

1. Enter a "complex" prompt ("Write me a python implementation of Dijstra's algorithm" here)
2. Interrupt the model during generation ("stop" here)