is friendlier towards reference cycles.  This change is not normally
visible since an optimized C implementation (_thread._local) is used
instead.