Public constructors for GamModelParameters

[najeeb-kazmi]

Part of #1698

Right now, in BinaryClassificationGamModelParameters and RegressionGamModelParameters we have constructors that look like the following;

internal BinaryClassificationGamModelParameters(IHostEnvironment env, int inputLength, Dataset trainset,
            double meanEffect, double[][] binEffects, int[] featureMap)
            : base(env, LoaderSignature, inputLength, trainset, meanEffect, binEffects, featureMap) { }

We want to make public constructors for all ModelParameters (previously predictors). However, for GAMs, this requires the user to pass in a training set, which is a Dataset object in Microsoft.ML.FastTree.Internal. This is not something we should expect the end user of the API to be aware of, so the constructor was not made public. Currently, Dataset is needed in GamModelParametersBase to map features to flocks and get bin upper bounds. We should rewrite GAMs so that this information is passed on from the Trainer rather than be computed when constructing the GamModelParameters.

[rogancarr]

A lot of the information that the trainer passes to the GAMPredictor is internal training information. I'd like to separate this out so that GAM preprocesses this. We would also get a GAMPredictor object that we could then use with different methods of training, which is fun.

[glebuk]

@najeeb-kazmi - please verify that is done and close.