Take the Bull by the Horns: Hard Sample-Reweighted Continual Training Improves LLM Generalization