MNT Support numpy2

[adrinjalali]

fixes #425

This adds support for numpy2. In the process, I also found some other issues which I fixed. Some of them were in tests for model cards, and I think the changed code is nicer since it leaves the titles of the sections unchanged.

[adrinjalali]

CI is green. Except that a single line is not tested since we're testing a recent numpy<2 here.

I'll have another PR to test minimum dependencies in CI, but for now this fixes urgent issues and I think we should merge and release.

cc @BenjaminBossan

(FYI, numpy issues on the skops side @TamaraAtanasoska)

[BenjaminBossan]

Thanks for updating skops to work with numpy v2 and for adding the other improvements.

[BenjaminBossan]

Honestly, I can't quite remember how it worked: Would adding an item here affect compatibility? I guess it's backwards compatible as witnessed by the test but not forwards compatible? Also, is this a change for numpy v2 or was it an oversight of not having this?

[adrinjalali]

So in numpy~=1 we have:

>>> np.random.default_rng().__reduce__()
(<function __generator_ctor at 0x777841d798a0>, ('PCG64', <function __bit_generator_ctor at 0x777841d3cc20>), {'bit_generator': 'PCG64', 'state': {'state': 220935942547292961595576198891696034818, 'inc': 281084058663618135801235667263986488477}, 'has_uint32': 0, 'uinteger': 0})

while numpy=2 does this:

>>> np.random.default_rng().__reduce__()
(<function __generator_ctor at 0x778c98b64b80>, (<numpy.random._pcg64.PCG64 object at 0x778c99543480>,), None)

and that means our check fails, and while fixing the check going recursively in what's in the output of __reduce__, I noticed we're not saving the seed sequence's state. It doesn't seem to have an affect in the next few generated numbers as I checked, but it's something which can be stored and loaded, so I added it.