Consider sklearn way of fitting

Brought up by @gwarmstrong

In sklearn the `fit` function is where the data is input rather than the model constructor. Would be more intuitive to users familiar with sklearn but could require some rethinking of keeping feature names, sample names, etc.

Start by reviewing the current model constructor and fit interface, then compare them with the scikit-learn fitting convention described in the issue. Trace how feature names and sample names are currently retained. Done requires an agreed design for moving data input into fit without losing that metadata.

Written by the indexing model from the issue text.