TrivialEstimator class become ITransformer (Approach 2)

[artidoro]

I am a bit concerned about the fact that we have to .Fit(...) for non trainable transforms. I think it can make ML.NET look really bad! I worked on a solution for the problem...

I tried three approaches:

1. Create a new interface that combines IEstimator and ITransformer:ITrivialEstiamtor : IEstimator , ITransformer
2. Make the trivial estimator class implement ITransformer:TrivialEstimator : IEstimator, ITransformer
3. Simply add a method .Transform(IDataView input) in  the TrivialEstimator class.

For each one of these the difficult part is making sure that the estimator/transformer chains work properly. So I implemented a TrivialEstiamtorChain that mimics EstimatorChain but on the class TrivialEstimator instead of the interface IEstimator.

For 1. and 2. I did not manage to make it work... When you .Append() two trivialestimators it will just have a difficult time to distinguish whether you want to create a new EstimatorChain or a TransformerChain. 

For 3. I managed to make it work, I think! The .Append() works and will switch to an EstimatorChain as soon as a trainable estimator is added to the pipeline.

Is it fine not having a class/interface that combines ITransformer and IEstimator?