In-context Autoencoder for Context Compression in a Large Language Model