Revisiting "Over-smoothing" in Deep GCNs