[Refactor] Code clean up for better readability

[garrett4wade]

Currently AReaL has too many levels of wrappers, which prevents potential users and contributors to read and understand the code.

We aim to clean up the code in two different levels:

1. the experiment launch procedure
2. the interaction between master and model workers

We don't have a concrete plan yet. I open this issue to share the current code structure and would like to discuss with the community how to do the refactor.

Readers can also check this amazing deepwiki page to understand the code structure.

Old documentation for ReaLHF: https://openpsi-project.github.io/ReaLHF/

The current launch procedure

The interaction between master and model workers

todo

[aryanraj2713]

@garrett4wade I am really interested in the work done, this issue looks like a perfect way to understand the repository, I believe we can divide this issue in multiple parts and I can work on refactoring