Large model class refactoring: Introduce ...AdapterModel classes

[calpt]

What does this PR do?

This PR performs a whole series of refactorings to...

... model classes:

• move all additionally introduced model classes out of the original modeling_X.py modules into separate modules in transformers.adapters.models
• move all mixins to transformers.adapters.mixins
• introduce XAdapterModel classes as replacement for XModelWithHeads classes to provide a clearly identified single go-to class for all adapter-related features.
• deprecate all XModelWithHeads classes
• integrate the prediction head definitions formerly in XModelHeadsMixin into XAdapterModel to simplify implementations

... tests:

• move all tests required for adapter-transformers to the new tests_adapters folder. It's therefore sufficient to only run tests in this folder for CI.
• refactor test classes such that each supported model has exactly one test module in tests_adapters that defines both the model test class (XAdapterModelTest) as well as all adapter test classes (XAdapterTest).

.... documentation:

• update model class docs to only show autodoc for XAdapterModel classes. All other class docs can be obtained from HF
• add a model overview page
• replace mentions of XModelWithHeads with XAdapterModel

... misc:

• add an additional __init__.py to allow all adapter-related classes to be imported via transformers.adapters
• update the adding_adapters_to_a_model guide

[JoPfeiff]

🥳

[hSterz]

Looks good