UniMem: Towards a Unified View of Long-Context Large Language Models