[ML] Support multinomial logistic regression in data frame analysis

[tveasey]

This issue tracks introducing multinomial logistic regression. The main complication of this is that the loss function has multiple parameters and the forest needs to predict all of them. This breaks the assumption that we have a single prediction, gradient and curvature. The task therefore breaks down into some preliminary tasks independent of actually implementing the new loss function.

• Extend CLoss and CArgMinLoss to support multivalued predictions, gradients and a Hessian matrix
• Rework caching to store multivalued predictions, gradients and the Hessian. This affects how we index values in the data frame, the memory calculation, etc
• Rework the code to compute aggregate derivatives for candidate splits
• Rework the code to choose the best candidate split
• Extend CBoostedTreeNode to support multi-valued leaves
• Extend the inference model definition to support multi-valued leaves
• Extend feature importance to support multi-valued leaves
• Write the new loss function
• Wire in to classification api

[tveasey]

This is now complete and merged into 7.7.