Support Roberta on accelerate

[younesbelkada]

What does this PR do?

This PR adds Roberta model family support with accelerate ! This aims to support int8 quantization for these models.
Before merging I have notice few nits that I need to fix and discuss! I am unsure whether these fixes should be here or in accelerate.

1- If the model has the attribute _keys_to_ignore_on_save, it seems that it does not get properly initialized by accelerate (but I might be missing something here). AFAIK all models that has accelerate support for now, has at most the attribute _keys_to_ignore_on_load_missing but not _keys_to_ignore_on_save. Therefore when the base model gets saved in the accelerate test, these parameters does not get saved in the state_dict. I had to come up with a modification in the _load_pretrained_model function to randomly initialize these parameters since they're ignored by the _load_state_dict_into_meta_model. This post processing trick happens here.
2- Therefore I had to change the accelerate tests. Since the parameters that are assigned in the _keys_to_ignore_on_save are initialized randomly, I propose to check the logits compatibility between the base model and the accelerate model only for the attention outputs and not the lm_head output. These modifications happens here - maybe this modification could happen in the super class?
3- Last nit: in the accelerate tests, it is better to not override the variable inputs_dict since inside the main loop, we can switch from a xxxForMultipleChoice to a xxxForQuestionAnswering model. Therefore, this variable does not get modified by the class function _prepare_for_class since it gets modified only if the model is a MODEL_FOR_MULTIPLE_CHOICE model. and not reset to the correct inputs_dict afterwards.

Need also to fix the slow test for lilt model where I am having ValueError: embeddings.position_ids doesn't have any device set.

cc @sgugger

[younesbelkada]

These tests could go on the super class, but not sure

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[younesbelkada]

Closing in favor of #19906