Missing head_mask and decoder_head_mask arguments in encoder-decoder models

[stancld]

🚀 Feature request

Following the PRs #9569, #9634 and #9639, there are other encoder-decoder models, which either do not support head_mask and decoder_head_mask input arguments at all or can be only provided with a single head_mask argument used for head masking both in encoder and decoder. It would be, therefore, nice to make this feature uniform over all the decoder-models.

---

Models:

Model	Pytorch	TensorFlow	PR	Copy dependency
BERTGeneration	☑️	✖️	-	-
EncoderDecoderModel	☑️	✖️	-	-
FSMT	✅	✖️	#9819	-
LED	✅	☑️	PT - #9856 ; TF - #9988	-
ProphetNet	☑️	✖️	#9964	-
Longformer	✅	☑️	PT - #9856; TF - #9988	LED

Your contribution

I'm happy to add this feature in the following days, both for PyTorch and TensorFlow models. (Likely in shorter PRs in order not to create large, overwhelming PRs)

---

Reviewers: @patrickvonplaten, @jplu, @sgugger, @LysandreJik, @stas00 .