Trainability for Universal GNNs Through Surgical Randomness