Fix CI to use correct python versions

[aulemahal]

No description provided.

[aulemahal]

Well, the CI is fixed at least...

[aulemahal]

@huard, I believe this is a case where following the SPEC0 is easier. According to SPEC 0, we are allowed to drop support for python 3.9, 3.10 and 3.11. I'd say we keep 3.11, but drop the two others. Instead of fixing the issues.

There is also a forgotten scipy call to fix. I don't understand why scipy is installed here, on all builds, but not in the xesmf-feedstock build. Not sure how to make a CI build without it as it is not requested anywhere...

[huard]

The mask creation code got a lot simpler. It is exercised by the test suite I assume ?

[aulemahal]

It is in the tests!

The last error here is sadly significative. For one case, the spatial averager, returns a wrong result. It happens with the "segmentized" version of the polygon, not with the "raw" one. And it only happens with 3.11 and esmf 8.4. The shapely version is the same in all envs, so my guess is that this comes from ESMpy/ESMF.

ESMF 8.4 is the last available version for windows so I don't think we should drop it. I'll look into the issue in more details.

@sol1105 Sorry about the delay...

[aulemahal]

I have not understood the issue, but I have avoided it. When we segmentize the test polygons with exactly 1, the first test polygons yields wrong weights, only in ESMF <= 8.4.2. The changelog for 8.5 says something about "concave" mesh elements, maybe its related ? I can't test with ESMF 8.5 as it is not available on conda-forge. But the issue isn't raised wityh 8.6 and above.

I avoided the issue by segmentizing with 0.99 ....

@huard I will release yet another patch after merging, I already updated the changelog.