Added log loss classification metric

[svarqq]

I based the implementation on sklearn's, with these main differences:

• Dropped labels as an optional argument, since it's mostly useful for string labels. Note as a result, the class index or class number of a value in y_true is always given by its ordered ranking over all y_true values (in the language of sklearn, the class index is always inferred). For example, if y_true = [4, 0, 9, 4, 9] the indices would be [1, 0, 2, 1, 2].
• No clipping of y_prob values to interval (eps, 1-eps). That means log(0) can be evaluated to -Inf which would immediately cause a return of Inf for the function. I chose to keep this possibility because it immediately signals either that the client got inputs wrong, or that they should retrain their model/forecaster. Plus, I did try clipping to the eps interval using Nx.Constants.epsilon/2 for the underlying floats in the tensor, and still got Inf's returned.

[josevalim]

💚 💙 💜 💛 ❤️