Loss of plasticity in deep continual learning