fairseq (0.10.2) replace-unk broken because of wrong data format

[jm-glowienke]

Alignments are given as tuple, which causes the following piece of code to break in fairseq/utils.py:

src_token = src_tokens[alignment[i]]

Cannot use tuple to look up element of list.

TODO:

• [x] fix the code to use the first element of the alignment, source index
• [x] open PR with fairseq
• [x] change use of fairseq to github installation, first from Marc's own github, when merged switch to fairseq/master
• [x] add feature to use original input sequence for replacement of unk, when word is also not included in source dictionary
• [x] fix tokenization for inputs to split word and sentence pieces to be used for replacement