Deprecate use of global PRNG in strategies

[Zac-HD]

Fixes #3810.

[jobh]

Looks good!

An idea: What if we check more randomness sources, but for performance reasons only occasionally? For example every nth draw, or random spot checks with p<0.1?

I'm saying this because I'd typically reach to np.random instead of math.random myself.

[jobh]

Maybe consider factoring this out as

with checked_random_state():
    result = self.conjecture_data.draw(strategy, observe_as=f"generate:{desc}")
    ...

?

That would avoid the duplication w/control.py, plus improve readability in these central functions

[Zac-HD]

Yeah, I like it - and we can check a "drew from st.random_module() flag there too.

[Zac-HD]

An idea: What if we check more randomness sources, but for performance reasons only occasionally? For example every nth draw, or random spot checks with p<0.1? I'm saying this because I'd typically reach to np.random instead of math.random myself.

Unfortunately Hypothesis will then complain about Flaky errors when it goes to shrink such things.

Maybe we check that less frequently? Or only if you set a flag? Actually I should check the perf cost of doing it every draw, it might not be that bad with if "numpy" in sys.modules:...

[jobh]

Unfortunately Hypothesis will then complain about Flaky errors when it goes to shrink such things.

Ah, yes, of course. We would have to flag the test and warn when done, I guess. No need to go that far in this PR IMO, it can be added later.