Baseline LearnerLCE that fits a straight line through the most recent
window per-batch training performances and linearly extrapolates it to
the requested test batches.
This is a stronger "no-curvature" baseline than the constant predictors of LearnerLCEFeatureless and matches the "rolling slope" baseline from the celecx research plan. The line is fit on the task's lce_link scale.
When predict_type = "se", the standard errors come from the ordinary
least-squares fit: se_epistemic is the SD of the fitted mean (so it grows
as the requested batch moves away from the window, reflecting extrapolation
uncertainty) and se adds the residual variance back to form the total
predictive SD. Both are on the link scale and are NA when fewer than three
window batches make the residual variance undefined.
Creates a new instance of this learner.