Training Statistics for linear learners.

[sfilipi]

When trying to split OLS as a separate learner, on PR #611, we noticed that LogisticRegression has a dependency on MKL to generate training stats.

Since we don't want the core ML.Net package to have dependencies on other large packages, i am removing the stats generation through PR #611 and should introduce a Utility, in the AdditionalLearners package, (or the package that has a reference to Mkl) and produce those stats post training.

[montebhoover]

This will fix #1006.

[sfilipi]

@GalOshri @TomFinley @shauheen I am thinking to add another LogisticRegression trainer in the HAL learner package, that computes the training stats wit SDV making use of MKL.

The ComputeTrainingStats method is called at the end of training from the base class.

The other alternative would be, as Tom suggested, to create an utility method that takes all the state necessary to re-compute the training stats. This would be limited to programmatic use, and to me feels less usable than introducing an alternative learner.

Open WIP #1392. we can discuss there as well.