Large refactoring, many new features, lots of changes

[Kipok]

Sorry for the huge PR.. Will merge without review since changes are too large - we can fix things in follow ups.

Main things that have changed are the following:

1. A lot of refactoring on core classes. Now the server code is fully isolated and can be used with a straightforward interface and no code execution if that's not required (see inference docs).
2. Support for retrieval-based prompts and associated refactoring on the prompt class.
3. Trtllm server is completely new and ~2x faster.
4. A lot of new tests, most of which are not run in CI, but can be triggered locally (see tests/gpu-tests/run.sh)
5. Direct support for nemotron chat format for both inference and training
6. Support for multinode servers
7. Some renaming for various flags / scripts to better align with new code structure
8. Some improvements on math evaluation
9. All packages are upgraded to latest versions

Some known broken things:

1. Documentation with --help
2. OpenAI model
3. Data explorer tool