Make tp_watch thread safe

  assign_version_tag in PyType_Watch is protected by the lock, but
  tp_watch is also assigned in PyType_Unwatch.  This makes both use
  an atomic operation so watch/unwatch requests won't get lost.

Objects/typeobject.c

@@ -772,8 +772,8 @@ PyType_Watch(int watcher_id, PyObject* obj)
     // ensure we will get a callback on the next modification
     Py_BEGIN_CRITICAL_SECTION(type);
     assign_version_tag(interp, type);
+    _Py_atomic_or_uint8(&type->tp_watched, 1 << watcher_id);
     Py_END_CRITICAL_SECTION();
-    type->tp_watched |= (1 << watcher_id);
     return 0;
 }
 
@@ -789,7 +789,7 @@ PyType_Unwatch(int watcher_id, PyObject* obj)
     if (validate_watcher_id(interp, watcher_id)) {
         return -1;
     }
-    type->tp_watched &= ~(1 << watcher_id);
+    _Py_atomic_and_uint8(&type->tp_watched, ~(1 << watcher_id));
     return 0;
 }