Scikitlearn clustering methods cleanup

I think these methods could do with a review. Here are a few things that look like issues to me:

• Some models do not implement a transform method (presumably because python scikit-learn does not) but could, no? Example: DBSCAN

• Recall that MLJ makes a distinction between report and fitted_params: the latter is for the learned parameters (in this case what is needed to assign a new observation to a class), and everything else goes in report. It seems that in the scikitlearn clustering wraps everything is just lumped into fitted_params. In particular this has led to inconsistency with the Clustering.jl models KMeans and KMedoids (which separate things correctly, as far as I can tell).

cc: @tlienart

The issue concerns the scikit-learn clustering wrappers, especially DBSCAN, and compares their transform behavior and fitted_params/report split with Clustering.jl's KMeans and KMedoids. Start by inspecting those wrappers and the referenced model interfaces; done means the clustering methods consistently expose supported transforms and separate learned parameters from report information.