[AutoML] Extend use of AutoML inferrence workflow for non-experiment types

I've got a desire here to use some of the transformer inference and other cool things from AutoML but obviously doesn't fit as an experiment without a proper estimator. I've mocked up adding anomaly detection as an experiment base by faking the estimator. Looks like too much of AutoML is internal for me to extend it like this. Any thoughts on how I should go about achieving the same results? Maybe I should setup a pull request to change the appropriate protection levels for the items needed to implement the experimentbase class?