[WIP] fit size reduction

[jeremyrcoyle]

WIP PR that adds Lrnr_base$reduce_fit to drop fit_object components not necessary for prediction, reducing memory demands.

In response to https://github.com/tlverse/sl3/issues/12

• Implements Lrnr_base$reduce_fit, which drops out fit_object components
• Adds private$.fit_can_remove, which takes a character vector of fit_object components not necessary for prediction
• Adds options("sl3.reduce_fit") If true, we automatically call reduce_fit as part of the training process
• Adds test-reduce-fit.R, tests fit reduction for a few learners that I've verified so far