Teacher training configuration improvements

[ZJaume]

I think the teacher-big task configuration used here

translations/pipeline/train/configs/model/teacher.yml

Line 12 in e714bb7

task: transformer-big

and here
can be optimized.

Regarding the speed of training:

• beam-size to 4 should be enough for transformer-big models.
• valid-mini-batch to 8 is a bit low, could be set to 32 or 64.

Regarding quality:

• max-length is set to 100, which is pretty low on my opinion (I typically use 400). Specially if we are including sentences from certain EU corpora from OPUS that have very long lines and segments from HPLT in backtranslation (remember that HPLT does not have sentences splitted, they are in the corpus just as they appear in the original website). The valid-max-length is set to 300, which is ok, but max-length to 100 is causing all the training sentences over 100 tokens to be omitted, so the model is not learning from them (unless I'm missing a third configuration file in the pipeline).
• I've been using swish always with no issues, but maybe there's no difference in using relu.

[eu9ene]

Related to #89