Verify that we can use NIM containers on slurm and locally in a convenient way

[Kipok]

Our code is already modularized enough to allow using NIM containers, but we never verified that and so some small adjustments might need to be made. Here are the instructions on how to get started with NIMs https://docs.nvidia.com/nim/large-language-models/latest/getting-started.html.

From our side what we need to do is to have a very streamlined interface to this, so that we just provide one of those containers as an input, either setting in a config and then adding server_type=nim (?) or directly as an argument, whichever is more convenient. This needs to work both locally and on slurm and we should ensure that it's easy to setup the proper caching of the model that container is going to download