Homogenize sign of translation coregistration and add affine tests

[rhugonnet]

This PR fixes and homogenizes the sign of translation parameters in coregistrations, and adds consistent tests mirrored from test_biascorr to ensure that affine coregistrations behave as intended for all types of input (raster-raster, raster-point, and point-raster).

In short: the new tests introduce synthetic shifts, then check that estimated coregistration shifts are exactly the same as those introduced within 1% (10% for ICP), and that applying the coregistration then corrects more than 99% of the variance in the initial elevation difference (95% for ICP). This ensures that the direction of coregistration is the right one, and that there is no other errors introduced in the process.

More background

Before #530, the _apply_rst of different method such as NuthKaab and ICP relied on different code to apply the translation. In some occasions, it called the same apply_matrix function, but even then it relied on a different to_matrix_func to derive the matrix despite depending on the same parameters (shift_x, shift_y, shift_z).

This meant that the signs of shifts stored in shift_x, shift_y, and shift_y could be inconsistent between methods. Since #530 relies on the same apply_matrix everywhere, this resulted in a erroneous coregistration direction for certain input types (point or raster) and certain method, which weren't capture by the older tests.

Additional fixes

Now that ICP is tested for all combinations of raster-raster, point-raster and raster-point with a subsample, we can indeed conclude that #422 and #423 indeed came from the prefilter issue of our old geoutils.interp_points that was removing data by default, which caused issues for point runs, see GlacioHack/geoutils#484 for details.

Resolves #422
Resolves #423
Resolves #485
Resolves #586

[rhugonnet]

@adehecq Merging this to use it to finalize #502, as it's mostly tests and your comments can easily be accounted for in a separate PR!