Speed up sampling of raster values

[dnomadb]

Most time for grabbing raster values for facet geometry is spent here: https://github.com/mapbox/make-surface/blob/triangle-indexing/makesurface/scripts/fill_facets.py#L57

To help with https://github.com/mapbox/weather-data/issues/71:

• Explore options for speeding up this method (rasterizing then extracting). Existing method iterates through each unique value and:
  1. Selects those features "burned" into raster with numpy.where
  2. Calculates the mean of these indices.
• Test if simple centroid extraction (very likely faster) would provide acceptable results.

cc: @ian29 @andreasviglakis @rclark @sgillies

[dnomadb]

Centroid extraction (pulling data values from a single pixel within the finest triangles) is looking to be a good bet.

Speed Improvements

As this would allow us to skip rasterization, all np.where, and np.mean calculations, it is much faster - on the order of ~100x (anecdotal observation)

Accuracy

Testing for this over the weekend. Based on the amount of upsampling we are doing before sampling, I would think that the difference will be negligible, but this is yet to be determined.

[dnomadb]

This method is much faster than (rasterio.features.rasterize --> np.where --> np.mean), and preliminary results look ok. Going to use it and punt on accuracy testing for now, as the speed increase is significant.