Refactor Roberta using modular

[Cyrilvallez]

What does this PR do?

Refactor Roberta using modular. It also improves the modular converter in the following ways:

• If any decorators are used on top of classes/methods in the modular_xxx.py, those are used instead of the decorators in the original class
• If any variables used in the decorators are missing, automatically look for them in the visited files
• Automatically removes unused variables that may be written as part of the decorators dependencies, but that are not used if the decorators were overwritten
• Fix the replace_call_to_super which was not behaving correctly if any statements appeared before the super() in the modular_xxx.py

All modified files come from the copied from Roberta.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[Cyrilvallez]

cc @ArthurZucker, could you review the changes please? 🤗

[Cyrilvallez]

Ok, we now have a better dependency mapping when replacing class nodes in ModularConverterTransformer:

• we only look for (recursive) dependencies of the new node (after the call to replace_call_to_super). This avoids adding wrong dependencies, and thus removes the need for post-cleaning of classes/functions
• This is important to not rely on post-cleaning for functions/classes because since imports will be taken care of in the post-cleaning, we could otherwise still keep some of them wrongly (if a "to-be-removed" function/class is the only reason to keep an import)
• if we have no inheritance from a model-specific class, only check the files for functions imports, and add them and their dependencies if needed

[Cyrilvallez]

IMO the logic in leave_Class in ModularConverterTransformer is much simpler now, let me know what you think I may be biased 😅🤗