In-Context Learning State Vector with Inner and Momentum Optimization