Idea for a two step preprocessing pipeline

[perib]

Currently, the preprocessing pipeline is applied to the entire training set before the evolution of pipelines. This is fine for things like one hot encoding.

We may want some parts of preprocessing to be trained per fold of cross-validation, such as iterative imputer. However, this is expensive and should be done only once.

This is something we may consider implementing into TPOT2. However, it is also something users can implement themselves as a custom objective function as well.