Number support

This just adds some basic support for voice recognition for numbers. Ideally, we'd be able to pull this out of nlp_compromise, but best I can tell they don't export their numbers grammar. As such, I've just reimplemented it roughly in JSGF. Everything seems to work, although something along the pipeline is overly eager and routinely cuts off long utterances. e.g., "two thousand three hundred and forty two" will often (but not always) come through as "two thousand three".

Clearly I messed up the grammar a couple of times doing this, thus the "show me my broken grammar" patch.