Allow InternVideo2/multi_modality to be pip installed

[rongou]

We are trying to use the InternVideo2 multi_modality model, but right now it's quite painful, we have to git clone the repo, fix the imports, and use a custom pyproject.toml to pip install it.

This change makes it a little easier. Once merged, you can run:

pip install \
"git+https://github.com/OpenGVLab/InternVideo.git@main#egg=internvideo2_multi_modality&subdirectory=InternVideo2/multi_modality/"

to install it. For the two flash_attn modules, provide more pip workers:

MAX_JOBS=24 pip install \
"git+https://github.com/OpenGVLab/InternVideo.git@main#egg=internvideo2_multi_modality[extra-git-deps]&subdirectory=InternVideo2/multi_modality/"