gh-141510: Add can_modify_dict() in dictobject.c

[vstinner]

can_modify_dict() is stricter than ASSERT_DICT_LOCKED() for frozendict. It uses PyUnstable_Object_IsUniquelyReferenced() which matters for free-threaded builds.

Replace anydict_setitem_take2() with setitem_take2_lock_held(). It's no longer useful to have two functions.

• Issue: PEP 814: Add built-in frozendict type #141510

[corona10]

Just to share my thoughts:

1. When we implement an xxx_lock_held function, I believe the goal is to ensure that developers declare critical sections properly through assertions. This change could cause confusion if someone starts calling xxx_lock_held, since it may require ASSERT_DICT_LOCKED for mutable collections but not for immutable ones.
2. My personal view is that assertions should report consistent results whenever possible. As I understand it, PyUnstable_Object_IsUniquelyReferenced does not guarantee this. Its result depends on the runtime situation, which means we could miss buggy scenarios until they are triggered in CI.

It does not mean -1 on this change but prefer to listen other core devs opinion.

[corona10]

I changed my mind after DM with Victor.
Since the current assertion is only meaningful when the refcount is not 1.

cpython/Include/internal/pycore_critical_section.h

Lines 55 to 58 in c6a142f

	# define _Py_CRITICAL_SECTION_ASSERT_OBJECT_LOCKED(op) \
	if (Py_REFCNT(op) != 1) { \
	_PyCriticalSection_AssertHeldObj(_PyObject_CAST(op)); \
	}

[vstinner]

This change does multiple things:

• It prepares dictobject.c to access frozendict without locking. @corona10 already made the first change for that in _PyDict_FromKeys() (commit d1b541b).
• Check for bugs: we must not modify a frozendict if its ref count is greater than 1 (since frozendicts are immutable!). ASSERT_DICT_LOCKED() check is not the right function to check that.

[vstinner]

Merged. Thanks for your review @corona10.