Out-Of-Context Prompting Boosts Fairness and Robustness in Large Language Model Predictions