Update YOLO Preprocessing

[kimichenn]

closes #277

Update some yolo stuff to support the new structure of the model.

Notable changes:

• Detection threshold bumped to 0.35 (from 0.05)
• Threshold is now a setting in config jsons instead of scattered everywhere (the only exception is unit/integration tests, as I assume we want to mess with this number often)
• Added NMS to deduplicate bboxes
• Changed default image size from 640x640 to 1024x1024 (I tried both image sizes for training/inference, and the 1024 works better so we'll stick with that)

The model can be found here.

I don't foresee us using the old model, so this can be merged regardless of whether I make future updates to the model (the architecture should be the same), with a big caveat, as detailed below. I think we'll just stick to finetuned YOLO this year.

The caveat: it is important to note that this model does NOT work for any not-stolen images (this might change after we update not-stolen to only use human and tent images). This means that we cannot run the full sitl suite unless you manually add detection yourself in the reports page.

[AskewParity]

I haven't actually looked to test this model on sitl, but the detection images you sent me look very promising.