Stop saving copies of the dataset in fitted estimators

[EgorKraevTransferwise]

Right now, we appear to be saving copies of the training dataset in each estimator being fitted, leading to huge memory requirements.
Is your feature request related to a problem? Please describe.
A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]

Describe the solution you'd like
A clear and concise description of what you want to happen.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.