LoCoCo: Dropping In Convolutions for Long Context Compression