Training ReLU networks to high uniform accuracy is intractable