[ML] Increase the maximum number of classes we will support #2246
Description
Because we compute the Hessian of the cross-entropy, for classification we don't scale well to many classes and have limited the maximum number to 30.
We can likely already increase that limit as a result of performance enhancements. However, to increase it significantly we need to think about how to tackle this bottleneck. A reasonable option to explore is identifying classes which are well separated in feature space. For any given class pair (i,j) for which there aren't regions where you are choosing between i and j then one could safely zero {H(i,j), H(j,i)}. This would in effect create block diagonal Hessians and we can bound the maximum block size as we do now.