Safety Pretraining: Toward the Next Generation of Safe AI