Fine-tuning Reinforcement Learning Models is Secretly a Forgetting Mitigation Problem