[seq2seq] correctly handle mt5

[stas00]

This PR fixes seq2seq/utils.py to handle mt5 like it does t5.

Ideally there should be a test, which would require creating a tiny model for mt5, but I'm being told this code is going away anyway, so there is no point investing energy into it.

Fixes: #9865

@patil-suraj, @sgugger

[sgugger]

LGTM as a quick fix. @patil-suraj when you port this to the new run_seq2seq, it would be great to try to find a way to make this not use any special code for a given model (but that's out of scope of this PR).

[patil-suraj]

LGTM! Thanks Stas for fixing this!

[patil-suraj]

when you port this to the new run_seq2seq, it would be great to try to find a way to make this not use any special code for a given model

I'm working on it in #9844, it's not finished though. We might need to add get_input_embeddings and get_pos_embeddings methods to every s2s model, to avoid special cases.

[sgugger]

If we need to add some methods to deal with the special cases, I would prefer it (otherwise the script might fail with new seq2seq models).