Foreshadow upgrade cleaned up (#206)

* Upgrade SimpleImputer module path and refactor existing code and test

* Using column transformers to replace the parallel processor in the intent mapper

* Adding cleaner into the picture. It turns out that if we call fit on a pipeline, it will in turn call fit_transform on each step except for the last step. For the last step, it will call fit instead.

* Adding back feature summarizer. The pandas wrapper is most likely to wrap over the output of the transform method in each Preparestep. Otherwise we have to do a conversion in every step

* Need to override the stacking method in the ColumnTransformer class to enable DataFrame as a return value. This has impact on all PrepareSteps.

* Update the data exporter

* Turn off a unit test comparing predict_prob arrays with long decimal points while the scores are already close.

* Fix the index misalignment by set original index on the nan dataframe in drop cleaner

* Some small changes on console, random_search (not used), parallel processor (may be deprecated)

* Adding flattener, its test and changes due to flake8 failure

* Update unit tests for intent type, foreshadow and na filler to improve code coverage and consistency of accessing test data

* Refactoring of code to improve code coverage. 

* Remove parallel processor and DynamicPipeline

* Remove serializer and deserializer as we are not using it to save the trained model.