Optimize Natural Neighbor

[dopplershift]

Now that we have a pure Python implementation of Natural Neighbor interpolation, we need to do some work to make it fast enough for big grids. The options I see are Numba and Cython, and I lean towards Numba.
Pro Numba:

• No need to ship compiled packages
• Active work to continue to optimize stuff
• No need to annotate source with anything more than a decorator

Con Numba:

• Not necessarily an easy dependency (though Conda helps). We could also ship a do-nothing decorator so Numba is optional
• Might need to refactor code to run better under Numba

I saw some really cool stuff from the Numba tutorial at Scipy 2016, so I'm going to give that a shot first.

[dopplershift]

If we use Cython, this set of notes may be useful.