On the Robustness of Average Losses for Partial-Label Learning