PERF/ENH: allow Generator in sampling methods

[mzeitlin11]

• closes BUG: random_state should permit numpy.random.Generator #38100
• tests added / passed
• Ensure all linting tests pass, see here for how to run them
• whatsnew entry

The perf gap is pretty crazy:

In [10]: rng = np.random.default_rng()

In [11]: ser = pd.Series(np.zeros(10 ** 6))

In [12]: %timeit ser.sample(n=10 ** 4)
13.8 ms ± 231 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

In [13]: %timeit ser.sample(n=10 ** 4, random_state=rng)
569 µs ± 8.11 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

Leaving #28440 open because I think the end goal should still be to make using these the default. But assuming that change isn't too easy since will be a breaking change?

[jreback]

looks fine a question.

[jreback]

umm why is BitGenerator not in the 2nd part?

[mzeitlin11]

BitGenerator isn't exposed until numpy >= 1.18. Have added a comment about this and removed the else clause (since that would mean we're not type checking anyway so the definition doesn't matter)

[jreback]

thanks @mzeitlin11

might need to slightly revist after we drop 1.17 (shortly)