On Symmetric Losses for Learning from Corrupted Labels