Fairness and Robustness in Invariant Learning: A Case Study in Toxicity Classification