make dimensionality reduction less experimental

[mastoffel]

we're currently allowing to do dimension reduction and provide any sckit-learn method. Many of these will fail (default is PCA, which works). We need to figure out which ones fail and how to make this feature more reliable.