Towards keras3

[rtavenar]

Keras3 is out and constitutes a significant API change.

Typically, our shapelets backend is not compatible with keras3 as such, so we need to either:

• keep things as such, and restrict keras version to be <3
• switch to keras3 and force the keras version to be >=3 in the future versions of tslearn

EDIT: keras3 is not available for Python <3.9, so this would require to force Python>=3.9 too for tslearn. I'm okay with it, personnally.

[rtavenar]

OK, it's not final yet, at the moment it cannot deal properly with time series of different lengths for the Shapelet model, but it's still a basis for future investigations.