declearn v2.6.1

Released: 08/10/2025

This is a subminor release that adds upper bounds to some dependencies (scikit-learn, jax, tensorflow).
It aims to prevent DecLearn code to break because of recent API changes in these dependencies.

In the secagg module, it also fixes quantization issues based on float values casting (with numpy >= 2.0).
Details :
In numpy 2.0, summing floatN values with an OOB value will result in inf values, hence making quantization faulty. This was caught in our unit tests for float16 inputs.
To fix this, backend checks have been added to determine the minimal float precision required for quantization operations and use either that or the input float precision.

GitLab release : https://gitlab.inria.fr/magnet/declearn/declearn2/-/releases/v2.6.1