Implement supervised training that doesn't require one million scripts

[ftyers]

Training should require two text files,

• SL side of parallel corpus
• TL side of parallel corpus

and four commands:

• Run SL side through the translation system until -biltrans
• Run the TL side through the monolingual system until -tagger
• Align both using something like eflomal

This will give three files, then a final lex-tools training using the above:

$ lrx-train biltrans target alignments > file.lrx