In-Context Former: Lightning-fast Compressing Context for Large Language Model