Multi-Turn LLM Conversations under the Least-Recently-Used Policy: Mean-Field Asymptotics