Added masked loss function from 10.7.5 to Seq2Seq

[nickmcgreivy]

Description of changes:

In 10.7, a masked loss function was added to the local Seq2Seq class using the decorator @d2l.add_to_class(Seq2Seq). However, this loss function was not added to the class d2l.Seq2Seq. This omission leads to incorrect training and validation errors in chapter 11.

In this pull request, I simply added the masked loss function to the d2l.Seq2Seq class.

By submitting this pull request, I confirm that you can use, modify,
copy, and redistribute this contribution, under the terms of your
choice.

[nickmcgreivy]

I also fixed a minor bug in the _tokenize function within the MTFraEng class (10.5). Previously the function returned max_examples+1 tokenized sentences. Now the function returns max_examples tokenized sentences.