ReSpec: Towards Optimizing Speculative Decoding in Reinforcement Learning Systems