[ICLR 2026] MeanCache: From Instantaneous to Average Velocity for Accelerating Flow Matching Inference