TextFeaturizingEstimator shouldn't use Argument class as main way of setting itself.

[Ivanidzo4ka]

machinelearning/src/Microsoft.ML.Transforms/Text/TextFeaturizingEstimator.cs

Line 141 in 5c00980

// REVIEW: expose them once sub-transforms are estimators.

Currently if I want to play with Word extractor and Char extractor I need to do that via

machinelearning/src/Microsoft.ML.Transforms/Text/TextFeaturizingEstimator.cs

Line 497 in 5c00980

public static IDataTransform Create(IHostEnvironment env, Arguments args, IDataView data)

which stops me from hiding IDataTransform interface.
We should have proper way to set it up via Setting object.

Related to #1995

[TomFinley]

Yes, we should. Though, that Create mostly exists for dependency injection via component factories, entry-points, and whatnot. So that can totally be private. We also can keep the associated fields on the settings object internal, so we can add a way to configure them via estimators (as described in the comments) elsewhere, I think? What do you think? Good, bad?

[artidoro]

From API standpoint, TextFeaturizingEstimator is fine once #2394 goes in. However, the entire class should be cleaned up.

[artidoro]

@Ivanidzo4ka after #2394, it seems that this issue has been resolved. If so I will close it.