[scikit-learn] How to not recalculate transformer in a Pipeline?

[Gael Varoquaux]

I use joblib.Memory for this purpose. I think that including a
meta-transformer that embeds a joblib.Memory would be a good addition to
scikit-learn.