Improve robustness of cebra model loading

[stes]

This PR

• Adds a test pipeline for numpy < 2. We want to support numpy<2 for ~ 1 extra year (discussed with @MMathisLab today), given its widespread use in other packages used together with CEBRA.
• Fixes a critical bug in the test_save_and_load which was skipped entirely previously due to a syntax issue. This revealed several issues in the loading logic
• As a result, the backend logic of CEBRA.load was improved: Instead of using three different backend choices, we now default to the (future proof) sklearn backend: The state dict of the model is saved, and the model is always constructed from the state dict. This follows the recent change in torch (from 2.6.0) to discontinue unsafe loads via pickle.
• Fixes an issue with model loading in python 3.14 by improving the cebra/registry.py module. When the @parameterize decorator is used, the class attributes are now properly passed to the wrapped class. A test has been added for this functionality
• The testing logic was broken on windows; a windows-compatible decorator for handling temp files was introduced.

[stes]

Just checked https://endoflife.date/numpy, it seems that <2 support actually ended a few months ago:

So I guess we could also ignore that particular test and force numpy > 2...

[stes]

Although e.g. Deeplabcut still has <2 requirement https://github.com/DeepLabCut/DeepLabCut/blob/85911cb83d315398ead65c1198e4991a73001834/setup.py#L68

[stes]

colab has 2.0.2

[MMathisLab]

🚀