Roadmap: parsnip support for sparse tibbles

[EmilHvitfeldt]

What we need:

• [ ] fit() to take sparse tibbles as data
• [ ] fit() to take {Matrix} sparse matrix as data
  • turn them into sparse tibbles early
• [ ] Have sparse tibbles turned into appropiate object before they are passed to engines fit function
  • {Matrix} sparse matrix if model supports it
  • back to normal tibble/matrix if not
• [ ] predict() to take sparse tibbles as data
• [ ] predict() to take {Matrix} sparse matrix as data
  • turn them into sparse tibbles early
• [ ] Have sparse tibbles turned into appropiate object before they are passed to engines predict function
  • {Matrix} sparse matrix if model supports it
  • back to normal tibble/matrix if not
• [ ] look into if we document which engines are sparse friendly
• [ ] special cases for some model types
  • {xgboost} with xgboost::xgb.DMatrix()
  • etc

I think we could use a option() of some kind to unit test that the data passed is passed around in a way that keeps the sparsity.

Adding all of this will give us

• standalone usage of sparse matrices in {parsnip}
• everything it needs to be able to work with the rest of {tidymodels} in regards to sparse tibbles