In Defense of the Unitary Scalarization for Deep Multi-Task Learning