Fix sampling in NuthKaab and set better default values for other classes

[rhugonnet]

This PR fixes:

• The sampling in NuthKaab that was done a bit too early, where NaNs introduced by slope/aspect could become problematic (somehow the tests were less sensitive to this).
• Classes such as Deramp that didn't have a default subsample value set for more reasonable computing times, now all changed to 5e5.
• The randomness of certain tests with a subsample, such as for Tilt.

Additionally, the previous way of installing the base environment for CI (concatenating YML dependencies without the Python version from a script get_yml_env_nopy) was failing on Windows when dependencies were pinned with a <, for example when I tried scipy<1.11.1 to ensure the randomness problem didn't come from there.
Now the script is renamed generate_yml_env_fixed_py, and generates a temporary YML file during CI to feed to conda, which parses all types of pinning without issues on Windows!

I'm leaving scipy<1.11.1, which also makes variogram modelling fail, will reactivate once they release 1.11.4: #430

[adehecq]

Seems good!
The changes in the tests are related to the fact that the automatically generated ddem has changed because of the change in the coregistration subsamplimg?

[adehecq]

Some other tests seem to fail though...

[rhugonnet]

Thanks! Yes it's the new subsampling that yields different values.

For the test not passing: it looks like the randomness of results is not fixed between different OS again... Joins #310. Will investigate...

[adehecq]

Haha, I like the many different attempts to make it work !! 🤣