Towards Understanding Quality Challenges of the Federated Learning for Neural Networks: A First Look from the Lens of Robustness