gh-135832: implement Py_DECREF specializations for Py_GIL_DISABLED build

[alex-semenyuk]

• Issue: Implement Py_DECREF specializations for Py_GIL_DISABLED build #135832

[colesbury]

I'm not sure this function is used anymore.

[colesbury]

This function is barely used now, but there's a similar function PyStackRef_CLOSE_SPECIALIZED.

There are two possible reasons to implement this as something that's not just Py_DECREF:

1. Performance, but it's not really clear that it's going to be any faster
2. To avoid _Py_Dealloc so that you can enable this assertion

   cpython/Objects/object.c

   Lines 3166 to 3170 in 6227662

   	#if !defined(Py_GIL_DISABLED) && !defined(Py_STACKREF_DEBUG)
   	/* This assertion doesn't hold for the free-threading build, as
   	* PyStackRef_CLOSE_SPECIALIZED is not implemented */
   	assert(tstate->current_frame == NULL || tstate->current_frame->stackpointer != NULL);
   	#endif

   , but that means you can't call _Py_MergeZeroLocalRefcount because that internally may call _Py_Dealloc

I think (2) would be difficult to achieve because of _Py_brc_queue_object.

[colesbury]

You can only safely deallocate if the reference count is zero and it's merged (i.e., new_shared is _Py_REF_MERGED)

[colesbury]

This is missing important parts of _Py_DecRefShared: if the shared reference count becomes negative, the object must be queued so that the owning thread merges the two reference count fields.