Switch to `numpy.random.Generator`

[flying-sheep]

Plan: SPEC7 (switch to rng: np.random.Generator | int args, phase out random_state: RandomState | int)

• rationale
• further context: leiden alg with igraph flavor causes out of bounds freezing #2969 (comment)

Implementation:

• Deprecate RandomState (using names only) #3372
• Implement transition tools and manually transistion in Add replace option to subsample and rename function to sample #943
• Decorator for functions that transition

[flying-sheep]

A big issue is that sklearn hasn’t done anything since 2020. That’s sad since they know how much this would benefit them (the following comment is a good read!): scikit-learn/scikit-learn#16988 (comment)

I think we only use it for PCA and (a RNG-less algorithm for) Neighbors, so if we can wrap things so the PCA keeps working, we could do it.

Then again, our PCA is already one of the most complicated parts of scanpy.