Use Mapnik geojson plugin (rather than CSV)

[springmeyer]

We should start targeting the Mapnik geojson native plugin, which like the CSV Plugin(#3) uses Mapnik's fast json parsing grammars. But unlike the CSV plugin we know the whole payload is geojson so there is less tokenization cost.

Specifically, after further optimization in Mapnik 3.x, profiling now shows the tokenization + quoting work to prepare to parse rows as geojson geometries is taking longer than parsing the actual geometries:

NOTE: this is with this geojson file with lots of features: https://github.com/mapbox/tilelive-overlay/blob/bench-async/test/data/complex.geojson

The catch is that the geojson plugin only supports featurecollections. So in the case of single geometries, they will need to be wrapped. But this should have minimal overhead and the case of multiple features will then go faster.

[springmeyer]

I'll add: in the case of single large geometries and not repeated features in a collection the CSV plugin is not likely much slower.

[tmcw]

👍 done in m3 branch.