Inference Optimization for MiMo v2.5: Pushing Hybrid SWA Efficiency to the Limit