Switch to Qwen3-VL

[gvzdv]

New Qwen vision model is out, and it's (on paper) better than the 2.5 that we use now.

I would like to try it out, which means we'll need to set aside time to swap models on Pegasus and test the new one.
There are several model variations (thinking/non-thinking, various sizes). I am considering a non-thinking (for speed) non-quantized 8B model — let's see if we can fit it on our GPU, especially if we can deprecate YOLO. Alternatives would be an 8-bit 8B model or an FP16 4B model.

[jeffbl]

Always a new model! Minor comment: be careful with "deprecate" which would be removing from the system forever. In this case, someone might want to run YOLO in the future, so we can simply remove it from our docker-compose.yml in favor of the LLM version, and get that memory back. In fact, we should definitely do so once we're running the LLM version.