Sleep-Like KV Cache Consolidation into Fast Weights Enables Long-Context Transformer Reasoning | HACKOBAR_