Quest: Add SAM Preprocessing for Videos

[gkielian]

One way LLMs can do video processing is via image segmentation, with tokens as labels for each entity in ascii.

We can really reduce context length required if doing this method.

Adding this recent repo for pre-processing, could then be used for next frame prediction, and with action in the loop (as demonstrated previously at Mistral SF hackathon):
https://github.com/facebookresearch/segment-anything-2