EvoP: Robust LLM Inference via Evolutionary Pruning