Output should match filter-rule-table.py, but filtering is faster. Some rough
timings:
That This
System A 0h 13m 0h 04m
System B 18h 03m 0h 51m
System A is WMT14, en-de, string-to-tree (32M rules, 3,000 test sentences)
System B is WMT14, cs-en, string-to-tree (293M rules, 13,071 test sentences)
This will eventually replace filter-rule-table.py. At the moment
it can only filter rule tables where the source-side is a STSG
fragment and when the test sentences have parse trees.