Simplify dependencies & split CPU/GPU backend Dockerfiles

[LouisAsanaka]

• Simplify dependencies into a single pyproject.toml file
• Streamlined local build through Docker / direct setup using uv
• Removed extra UniMERNet models to cut down image size (keeping the only base models)
• Added frontend Dockerfile
• Docker Compose is now working

[LouisAsanaka]

TODO: Download model in the workflow step before Docker build

[deluzhao]

A bit concerned about removing Omniparser. The file structure that I'm using here is not default and a pure re-install of Omniparser won't restore the file structure. I committed Omniparser because the non-downloaded model files are relatively lightweight. Why is Omniparser being considered unused? Is this for demo purposes?

[LouisAsanaka]

Is Omniparser still being used for the main pipeline? I think I mistook it being unused because it was commented out. I'll restore the directory.

Also, are only the weights of Omniparser used? We can move Omniparser code into the main source, and leave only model weights in models/

[deluzhao]

I think the configs are also used and I wouldn't remove the license just in case. There's probably some unused files, but this is largely just a restructured directory from Omniparser, so I didn't really try to touch individual files. I can look at reducing the files as a secondary priority after we complete the big action items.

[LouisAsanaka]

@deluzhao I've restored OmniParser and the relevant task.