Do we want to preprocess data within our algorithms?

[bellet]

Some algorithms include some data preprocessing: RCA and LMNN have a parameter to apply PCA to the data before fitting. I think a better practice (as done in sklearn) is to let the user apply preprocessing steps (such as normalization, dimensionality reduction) separately if needed, as it is more transparent.

If we do want to allow some preprocessing, when we should try to have at least a unified way to do this for all/most algorithms.