SIRI: Scaling Iterative Reinforcement Learning with Interleaved Compression