gh-153291: Fix data race in readline.get_completer() and get_pre_input_hook()

[harjothkhara]

readline.get_completer() reads state->completer and returns a new reference without taking the module critical section, while set_completer() swaps the same field under the critical section with Py_XSETREF. On the free-threaded build a concurrent getter can be incrementing the refcount of a hook the setter is simultaneously dropping. get_pre_input_hook() has the same problem with state->pre_input_hook — those are the only two hook fields with both a Py_XSETREF-based setter and a Python-level getter, so this covers the whole Python-visible getter surface.

The fix adds @critical_section to both getters, the same way gh-126895 did for the rest of the module.

Verified with a free-threaded TSAN build (--disable-gil --with-thread-sanitizer): the reproducer from the issue reports the race on main (the readline_get_completer read vs the set_hook write) and runs clean with this change. The new tests in test_readline.py trigger the same TSAN reports without the fix and pass with it.

Deliberately not included here: get_begidx()/get_endidx() and the C callback paths (on_completion() etc.) read module state without the critical section too, but their writer/reader sides live in readline callbacks (flex_complete(), on_hook()), where taking the module critical section is a design question (it would serialize completion callbacks against all other readline calls) — that's a follow-up, not a one-line annotation.

I used AI assistance while working on this change; I've reviewed it and can explain it.

• Issue: Data race in readline.get_completer() reads state->completer without the critical section #153291

[brijkapadia]

Is this test necessary? It didn't fail on main for me.

[harjothkhara]

Good point, that test just checks basic get/set behavior and passes on main, so I've dropped it. The free-threaded tests below are the ones that actually catch the bug.

[brijkapadia]

Also don't force push

[brijkapadia]

Instead of passing the barrier as an argument, you can just reference it directly:

Suggested change

	b.wait()
	barrier.wait()

You could then remove the arguments from the args parameter.
Same with the other test.

[brijkapadia]

I think the tests need to get moved to test/test_free_threading so that it runs with TSan.

[harjothkhara]

Moved them to Lib/test/test_free_threading/test_readline.py so the TSan job runs them. Verified they fail on main and pass with the fix.

[brijkapadia]

This other test probably also needs to run with TSan:

cpython/Lib/test/test_readline.py

Lines 449 to 463 in b2d8db1

	def test_free_threading(self):
	def completer_delims(b):
	b.wait()
	for _ in range(100):
	readline.get_completer_delims()
	readline.set_completer_delims(' \t\n`@#%^&*()=+[{]}\\|;:\'",<>?')
	readline.set_completer_delims(' \t\n`@#%^&*()=+[{]}\\|;:\'",<>?')
	readline.get_completer_delims()
	count = 40
	barrier = threading.Barrier(count)
	threads = [threading.Thread(target=completer_delims, args=(barrier,)) for _ in range(count)]
	with threading_helper.start_threads(threads):
	pass

It might be a good idea to move it in this PR although that's probably better answered by someone else