Make `_parse_sklearn` a public API for meta-estimator parsing

[khoover]

I've written up a few meta-estimator parsers for things like log1p target transforming (i.e. the TransformedTargetRegressor when the transform/inverse are fixed to np.log1p/np.expm1) or multi-model regression (i.e. multiple models running regression over the same input pipeline), and in both of them I've had to use _parse_sklearn within their parser functions to handle parsing the inner estimators. It would be nice to make this public, to support meta-estimators that can't be precisely expressed using combinations of Pipeline or FeatureUnion (e.g. an estimator that scales some regression by a classifier probability).

[xadupre]

Do you have a precise idea of the API you would like to use? A dummy example would help to implement it.

[khoover]

It would look roughly the same as how _parse_sklearn_pipeline uses it, except maybe it would add operations before/after the inner _parse_sklearn(scope, model.inner_estimator, ...). E.g. for the Log1pTargetTransformer, it would add an extra operator consuming the outputs of that parse call and later convert the operator into an exp and a minus-1.