handling NaNs in geolocation

[NiklasPhabian]

This MOD09 granule is somewhat botched:
https://ladsweb.modaps.eosdis.nasa.gov/archive/allData/6/MOD09/2002/299/MOD09.A2002299.0710.006.2015151173939.hdf

We have NaNs for the latitude and longitude at the beginning of the swath.

Unfortunately, we read those NaNs in without masking them; they end up being -999.
It then becomes really chaotic when we interpolate the 500 m geolocation.

Things we need to do:

1. Read in the mask/fill values correctly; we do that decently at STAREPandas: https://github.com/SpatioTemporal/STAREPandas/blob/master/starepandas/io/granules/modis.py#L77
2. Come up with a solution of adapting the STARE resolution to the neighborhood when a neighbor has a NaN value.
3. Come up with a solution to interpolate the 500 m resolution