Improving Representational Continuity via Continued Pretraining