Memory Consolidation Enables Long-Context Video Understanding