Beyond SFT: Reinforcement Learning for Safer Large Reasoning Models with Better Reasoning Ability