Compose Yourself: Average-Velocity Flow Matching for One-Step Speech Enhancement