Towards Robust Extractive Question Answering Models: Rethinking the Training Methodology