Investigating the Robustness of Sequential Recommender Systems Against Training Data Perturbations