[RAG] Fix Extra Positions Bug

[klshuster]

Patch description
--n-extra-positions can be used to expand the total number of positions used when encoding documents + input in pre-trained transformers. If we set this >0, then we are supposed to fill up all of end positions with the doc tokens, leaving the beginning positions for the input text.

In the current RAG implementation, if we specify --n-extra-positions M < --n-positions N, we encounter an issue where the documents were not properly truncated. This is found in the concat_docs_and_input function, where self.expanded_input_truncate is used to determine the maximum length of the documents. This is an overloaded attribute, hence why the bug was a bit subtle.

This bug was not previously caught, as in all of my experiments, --n-extra-positions > --n-positions.

Testing steps
Added a test to confirm that concat_docs_and_input respects the n_positions available in the model.

[mojtaba-komeili]

LGTM

[mojtaba-komeili]

nit: a general comment on self.n_extra_positions: In RagModel constructor we have assert opt['n_extra_positions'] >= 0. Yet there are line in code that we check for self.n_extra_positions <= 0. Can we remove the < check for clarity?

[jxmsML]

LGTM! thanks for fixing this subtle bug

[klshuster]

The failing test is fixed in #4395