• Closes Switch to numpy.random.Generator #3371
• Tests included or not required because:

• Release notes not necessary because:

The idea is to make the code behave as closely as possible to how it did, but make it read like modern code:

1. add decorator that converts random_state into rng argument (deduplicating them and using 0 by default)
2. add helpers that allow old behavior (e.g. for APIs that only take RandomState)
   • _FakeRandomGen.wrap_global and _if_legacy_apply_global to conditionally replace np.random.seed and other global-state-mutating functions
   • legacy_random_state to get back a random_state argument for APIs that don’t take rng, e.g. scikit-learn stuff and "random_state" in adata.uns[thing]["params"]
3. after this PR: make feat: presets #3653 change the default behavior when a preset is passed.

TODO:

• add the decorator
• add helpers for restarting (e.g. if 0 was passed, it’d be reused by the functions called in function body)
• for the functions that called it: re-add the np.random.seed calls (maybe if isinstance(rng, _FakeRandomGen): gen = legacy_numpy_gen(rng) or so?)
  • partially done, finish the work
• add spawn to _FakeRandomGen (that does nothing) and use spawn for a tree structure
  • ENH: Should there be an rng.clone() or similar? numpy/numpy#24086 (comment)
• figure out how to handle persistence for new RNG (random_state in adata.uns[thing]["params"])
• pass rng to neighbors transformer?