Sleep-Like KV Cache Consolidation into Fast Weights Improves Long-Context Transformer Performance | HACKOBAR_