gh-119213: Fix getargs.c to store state in InterpreterState...

[1st1]

...as opposed to storing it in PyRuntime. Storing it in PyRuntime
is fundametally wrong, as its state contains references to Python
objects. Those objects (tuples and strings) can (and will) be
picked by various subinterpreter clean up code, leaving PyRuntime
with broken pointers.

#119194 is a backport to 3.12

• Issue: getargs.c is Breaking Interpeter Isolation #119213

[erlend-aasland]

LGTM; perhaps it would make sense to init getargs state just after _PyGC_Init, but I'm not sure it matters; I'll leave that kind of nitpicking to Eric :)

[colesbury]

How does this work when the _PyArg_Parser instances are themselves declared as static variables (and are global to the process)? For example:

cpython/Modules/clinic/_struct.c.h

Lines 49 to 53 in 6f7dd0a

	static _PyArg_Parser _parser = {
	.keywords = _keywords,
	.fname = "Struct",
	.kwtuple = KWTUPLE,
	};

[vstinner]

@serhiy-storchaka: Would you mind to review this change?

@serhiy-storchaka wrote this API.

[ericsnowcurrently]

@colesbury, I'm looking into what's going on. Basically, the statically declared tuple is only for builtin modules. I have a solution that's different from Yury's but want to be sure it's correct before closing this one.

[ericsnowcurrently]

superseded by gh-119331