Reuse, Don't Retrain: A Recipe for Continued Pretraining of Language Models