How does Log Loss evaluate the probability predictions made by classification models? Why does Log Loss penalize confident incorrect predictions more heavily? How can developers use Log Loss to compare classification models and improve prediction reliability?