gh-117657: Fix data races in the method cache in free-threaded builds

[mpage]

These are technically data races, but I think they're benign (to the extent that that is actually possible). We update cache entries non-atomically:

cpython/Objects/typeobject.c

Lines 4974 to 4984 in a23fa33

	staticvoid
	update_cache(structtype_cache_entry*entry, PyObject*name, unsigned intversion_tag, PyObject*value)
	{
	entry->version=version_tag;
	entry->value=value; /* borrowed */
	assert(_PyASCIIObject_CAST(name)->hash!=-1);
	OBJECT_STAT_INC_COND(type_cache_collisions, entry->name!=Py_None&&entry->name!=name);
	// We're releasing this under the lock for simplicity sake because it's always a
	// exact unicode object or Py_None so it's safe to do so.
	Py_SETREF(entry->name, Py_NewRef(name));
	}

but read them atomically from another thread:

cpython/Objects/typeobject.c

Lines 5046 to 5051 in a23fa33

	if (_Py_atomic_load_uint32_relaxed(&entry->version) ==type->tp_version_tag&&
	_Py_atomic_load_ptr_relaxed(&entry->name) ==name) {
	assert(_PyType_HasFeature(type, Py_TPFLAGS_VALID_VERSION_TAG));
	OBJECT_STAT_INC_COND(type_cache_hits, !is_dunder_name(name));
	OBJECT_STAT_INC_COND(type_cache_dunder_hits, is_dunder_name(name));
	PyObject*value=_Py_atomic_load_ptr_relaxed(&entry->value);

and there's nothing that establishes a happens-before relationship between the reads and writes that I can see.

In practice I don't think this matters much. We care about always reading the entire entry atomically, and the sequence lock enforces that.

Sample races reported by TSAN:

• https://gist.github.com/mpage/8633aacbf11303bdfdbda2c0e644d43d
• https://gist.github.com/mpage/70d5758968c0225384a2ace448122d09
• https://gist.github.com/mpage/33436b62af0d2e688f36c4ecb9171dee

• Issue: Make TSAN tests pass with the GIL disabled in free-threaded builds #117657

[DinoV]

LGTM!