This has better multi-core performance and can run in persistent mode -- which
allows us to construct a test harness of all the parsers (which takes time) and
then just reuse the process for lots of different data.