fixed decoder_start_token_id for T5

[Ehsan-Jahanbakhsh]

in most t5 models pad_token is equal to decoder_start_token, so this wasn't a problem. but MADLAD-400 uses different decoder_start_token and pad_tokens and doesn't work. this fixes the problem

[vince62s]

I don't think that the config file ALWAYS has a decoder_start_token_id option set so maybe you need to use pad as default instead.

[vince62s]

you need to run black to fix the python style

[vince62s]

if you get some interesting results with MADLAD400 can you post something here: https://forum.opennmt.net/t/madlad-400-a-multilingual-and-document-level-large-audited-dataset-model/5487/2
thanks.