Fix issue with NaNs masks in Coreg methods

[adehecq]

This is a fix to #237.
During coreg.apply, a resampling is needed. When NaNs exist in the input DEMs, the NaNs also propagate and must be handled carefully. Currently, such resampling is done at 3 different places, and NaNs are not always caught properly.
This PR tries to fix this issue.

I also fixed a small issue in BlockwiseCoreg. Errors during fitting were caught, but not during apply.

[adehecq]

Note that BlockwiseCoreg is not very reliable with large gaps, but the current fix improved it a bit. See figure of elevation difference after Blockwise coregistration (See attached script inspired from test_blockwise_coreg_large_gaps): test_blockwise.txt

Before the PR

After the PR

The underlying NuthKaab method should probably be improved to be more robust with few points.

[adehecq]

Finally, this is another indication that Coreg methods should be thoroughly tested for datasets with large gaps (#140).

[rhugonnet]

Thanks for tackling this! It's good to have all these fixes incorporated right away.

I agree, we'll have to work on a testing scheme for outlier robustness! This can probably be done in the future by simulating data gaps/NaNs, which will also be super useful if we want to write a paper comparing the performance of Coreg methods, including robustness 😄 !

[rhugonnet]

Looks good!