Validate new apply_matrix() function further.

[erikmannerfelt]

As mentioned in #87, this function is quite convoluted and could potentially produce odd results in cases that tests don't yet properly cover.

@adehecq mentioned potential issues that may arise here:

xdem/xdem/coreg.py

Line 1127 in 8faaef1

if abs(matrix_diff[matrix_diff != matrix_diff[2, 3]].mean()) < 1e-4:

Would that not cause a problem if the matrix has other elements that are equal to element [2, 3]?
Also, should it not be strictly equal to 0? The 1e-4 value is somewhat arbitrary here...
In the worse case, we're just doing extra work for nothing.

and suggested:

Not the simplest way, but you could (make a copy,) set the unwanted value to 0 and calculate the mean.

It should be validated if this is a problem, and if so, we should try to fix it.

Other stress tests could be:

• Introduce test cases with many nans
• Validate the accuracy of sub-pixel shifts.
• Run a transformation back and forth many times and see if any bias drift happens.
• Try more extreme transformations (I would argue that 10 degrees of rotation is already extreme, but we could make it worse!)

[erikmannerfelt]

It seems like it struggles with many nans. This has to be checked further!

EDIT: Also @rhugonnet , you have my permission to say "I told you so!", but it seems like the mask dilation should be on by default, to not mess up the num-to-nan edges of a DEM.