Add Wav2Vec2 & Hubert ForSequenceClassification

[anton-l]

What does this PR do?

This adds a Hubert extension for sequence classification.
Ultimately this classification head should be compatible with s3prl UtteranceLevel implementation to support classification tasks from SUPERB, such as Keyword Spotting and transfer their pretrained models.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@patrickvonplaten @patil-suraj

[patrickvonplaten]

Yes! We have to discuss a bit with @Narsil how to best add those models to pipelines

[Narsil]

# Update this list for models that are not in any of the auto MODEL_XXX_MAPPING. Being in this list is an exception and
# should **not** be the rule.

Seems like the exception has grown quite a bit :)

[patrickvonplaten]

This PR looks to be in a very good shape already!
Before merging it would be great if we could:

• add a #Copied from Wav2Vec2 ... to the HuBERT code if it's 1-to-1 the same
• add one test per task for Wav2Vec2 as well
• add at least one model for each task to either https://huggingface.co/superb or facebook (let's check with others here)
• run eval of the models on the datasets to check which models should be normalized and which shouldn't and adapt configs accordingly

[anton-l]

Accuracy evaluation on SUPERB tasks:

• KS has uniform-length samples, so no padding
• ER has non-uniform padded batches
• SID is evaluated with batch_size=1 as in s3prl

Task	Model	normalize=True	normalize=False	Paper
KS	Wav2Vec2-base	0.9627	0.9643	0.9623
	Hubert-base	0.9669	0.9672	0.9630
ER	Wav2Vec2-base	0.5281	0.6258	0.6343
	Hubert-base	0.5502	0.6359	0.6492
SID	Wav2Vec2-base	0.7360	0.7518	0.7518
	Hubert-base	0.8071	0.8071	0.8142

So far normalize=False is always better, as expected (s3prl never used normalization during eval).
There's also some slight variation with the official results, but it's of the same magnitude as s3prl vs paper results.

[anton-l]

• Passed integration test for all 4 tasks on both models
• Added Copied from where possible (the script just inserts a full copy of W2V2.forward() before End copy, so I didn't use it there)
• Added dummy examples to forward() docs
• Moved the models to https://huggingface.co/superb

@patrickvonplaten everything should be ready to merge now :)

[patrickvonplaten]

Awesome job @anton-l ! Feel free to merge the PR whenever you want