pythongh-128691: Use deferred reference counting on _thread._local

After this change, the "thread_local_read" micro benchmark in
Tools/ftscalingbench/ftscalingbench.py now scales well when multiple
threads read from the same thread-local variable.

Modules/_threadmodule.c

@@ -1435,6 +1435,7 @@ local_new(PyTypeObject *type, PyObject *args, PyObject *kw)
     Py_DECREF(localsdict);
     Py_DECREF(sentinel_wr);
 
+    _PyObject_SetDeferredRefcount((PyObject *)self);
     return (PyObject *)self;
 
   err: