NeMo Skills requests

Thanks! Can we make a wrapper script that will make it simpler to use and support all the logic without users manually switching between code_execution / different number of gpus, etc. Or instead of a separate script, maybe an option to start_server --launch_chat_interface that when specified will add an extra command that starts the chat interface script? That way we don't need to have a separate docker and can reuse our existing functionality to launch containers. And if anyone wants to run inside existing docker, it will still work as long as cluster config is not specified. Plus this way we can specify all parameters in one place and not need to duplicate them in start_server and chat interface (e.g. to enable code execution).

Another thing I would add is an option to use predefined models, so people can just say --model=Nvidia/OpenMath-Nemotron-X and then we will fill in the correct config / template inside the gradio demo, so that it doesn't need to be specified. This will also make it simpler to avoid errors by specifying incorrect config (e.g. our kaggle model uses different config than non-kaggle ones for tir). Finally, the gradio should ideally support a switch between code execution and non-code execution with a button, not an argument as some of the models can be used in both modes (and we can disable it for the kaggle model, if we can recognize that it's used)

Originally posted by @Kipok in https://github.com/NVIDIA-NeMo/Skills/pull/464#pullrequestreview-2837737225

No files or tests are named. Start by locating the start_server entry point and the Gradio demo, then trace how container, model, and code-execution parameters are currently supplied. Done means a unified launch path, predefined model handling, and an in-interface code-execution switch without duplicated configuration.