Fix NaNs in "longyearbyen_ddem" example

[rhugonnet]

• Resolves Example "longyearbyen_ddem" contains NaNs #196
• Also fixes the use of Raster as an input of sample_empirical_variogram with a tiny bug on the ground sampling distance definition fixed, and tests updated
• Opened issue Preserve nodata in masked_array or Raster subclass during Coreg operations #225 to add the right behaviour directly in Coreg classes, so that there is no operation to perform for the user downstream

[rhugonnet]

For opening issue: we don't want to repeat this operation after every Coreg operation to preserve nodata along calculations:

aligned_raster.data[~np.isfinite(aligned_raster)] = reference_raster.nodata
diff = reference_raster - \
           gu.Raster.from_array(aligned_raster.data, transform=reference_raster.transform, crs=reference_raster.crs, nodata=reference_raster.nodata)

[rhugonnet]

It seems that GlacioHack/geoutils#257 fixed the issues that appeared in the tests.
We can merge!

[erikmannerfelt]

I'm afraid I don't really understand the fix here! You simplify the syntax and clean up the variogram test, which is great. How does this relate to the nans in the test?

[rhugonnet]

@erikmannerfelt The example bug fix is in the example. The rest is a side fix for using the Raster class directly as input, I added a line in the main description of the PR:

• Also fixes the use of Raster as an input of sample_empirical_variogram with a tiny bug on the ground sampling distance definition fixed, and tests updated

[erikmannerfelt]

Ah! Weirdly the example code diff didn't load on my GitHub. Now it does, and it looks good!

[adehecq]

This line and the few lines below could be simplified now that Coreg methods take Rasters as arguments and with the arithmetic overloading functions !