Pessimism Meets Invariance: Provably Efficient Offline Mean-Field Multi-Agent RL