Improved classification of subdivision/unit

[missinglink]

branched from stopword-classification, requires #179 be merged first.

This PR is an attempt at solving #178 with minimal changes.

However, it seems to solve a longstanding issue which was more widespread, namely having a space between HouseNumber and UnitNumber:

node bin/cli.js 10 A Main Street

# before
(0.98) ➜ [ { housenumber: '10' }, { street: 'A Main Street' } ]

# after
(0.86) ➜ [ { housenumber: '10 A' }, { street: 'Main Street' } ]

I'm going to have a think about this before merging, it feels 'too easy', I'm sure it comes with some pitfalls.

Closes #178

[missinglink]

These changes are required as the parser previously returned a single result for these queries, it now returns two, but the expected result is still the highest scoring.

[missinglink]

The improved unit parsing in this PR has caused some issues where in some cases the resulting parse, when using in pelias/api produces zero results:

https://pelias.github.io/compare/#/v1/autocomplete?text=Calle+Principal+20+B

vs.

https://pelias.github.io/compare/#/v1/autocomplete?text=Calle+Principal+20+Ba

[missinglink]

We may need to revert this is we can't come up with a workaround, one which comes to mind is to strip non-numeric tokens from the housenumber field in pelias/api before generating the subject field.

It feels like one step forward in support for units and then a step back 😢