Fixes issues with backward pass in LED/Longformer Self-attention

[aleSuglia]

Description

This PR fixes the computational graph created when computing the global attention scores in LED/Longformer Self-attention. The current implementation breaks the computational graph preventing the model from running the backward pass correctly in some cases. As explained in the dedicated issue, this problem arises in the current version of PyTorch (PyTorch 1.9) but not in one of the previous ones (PyTorch 1.7.1).
This PR simply clones the appropriate tensors in order to avoid the issue. clone() is a differentiable operation so there are no changes to the actual model behaviour.

The issue is carefully reproduced in the following Google Colab: https://colab.research.google.com/drive/13rKxs6Ype0kDEBlnywsGynE2zpzv2CR-?usp=sharing

Fixes #12613

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?

@ibeltagy @patrickvonplaten

[patrickvonplaten]

Thank you!