Fix usage of head masks by PT encoder-decoder models' generate() function

[stancld]

This PR adds missing arguments head_mask, decoder_head_mask and cross_attn_head_mask into prepare_inputs_for_generation function of PyTorch encoder-decoder models so that these args will be used during the generation when generate() function is called.

EDIT: Need to fix the new test for ProphetNet

---

Example

out = bart.generate(input_ids, ...)
tokenizer.decode(out[0], ...)

>>> 'The Eiffel Tower in Paris has been officially opened to the public.'

Behaviour before the PR:

out = bart.generate(input_ids, decoder_head_mask=decoder_head_mask, ...)
tokenizer.decode(out[0], ...)

- >>> 'The Eiffel Tower in Paris has been officially opened to the public.'

Behaviour after the PR:

out = bart.generate(input_ids, decoder_head_mask=decoder_head_mask, ...)
tokenizer.decode(out[0], ...)

+ >>> 'The Eiffel Tower in Paris has been officially opened to the public for the first time since it was completed in 1903.'

---

Reviewers: @patrickvonplaten

[patrickvonplaten]

Hey @stancld,

Thanks a lot for this contribution! Could we add one test to verify that generation works with head_mask for all encoder-decoder models?

I think it could be added to test_generation_utils.py

[stancld]

Hey @patrickvonplaten, I've added one test. At this moment, there are two little issues I'm gonna handle later today so that all encoder-decoder models will pass this new test.

[patrickvonplaten]

BTW, if Prophetnet doesn't work with head_masking + generate, I'm totally fine with leaving it out for ProphetNet - we could then just overwrite the test to not run for Prophenet. The model is very unique and also doesn't work fully at the moment in general

[stancld]

Hi @patrickvonplaten, sorry for being silent for a while as I've been a bit too busy. As you suggest, I skip the test for ProphetNetForConditionalGeneration model and now all the tests pass :)