[modular] Remove ambiguity in all calls to parent class methods + fix dependency graph

[Cyrilvallez]

What does this PR do?

A lot of modular files had wrong calls to parent's method (in order to skip unravelling the definition), making then non-pythonic files. This is now fixed, and the converter is much more robust on this.

Here are the new rules to make this process much more pythonic, such that modular files are correct python files:

• When we want to skip unravelling the parent's code (i.e. we do NOT want to call super), let's call the method from the actual class we would like the method to be used
• If it's a grand-parent of the inherited class, no need to add the base again in the MRO : e.g., if we inherit from LlamaMLP, and we want to call nn.Module.__init__(...), no need to re-add nn.Module as a new base, as it's already part of the MRO (LlamaMLP is a nn.Module, so it's the grand-parent)
• the converter will replace such class calls with super() to keep Python's best practices when the class called is one of the direct parents of the generated code

Overall, those rules are much more natural and pythonic, and clear the ambiguity that exists currently.

Also, fix a bug in how we were creating the dependency graphs (some models could be skipped due to non-exhaustive match)

[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.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: aimv2, aria, colqwen2, d_fine, deepseek_v2, deepseek_v3, deepseek_vl_hybrid, diffllama, doge, dpt, eomt, ernie4_5_moe, evolla, gemma3, gemma3n, glm4v

[Cyrilvallez]

This was not exhaustive enough so some models could be skipped from dependencies if we import simply their image_processor or such