Do Language Models Need Sleep? Offline Recurrence for Better Long-Context Reasoning