mlr3extralearners 1.6.0

Breaking Changes

• classif.priority_lasso, regr.priority_lasso, and surv.priority_lasso: the parameter sets were reduced to a focused subset of prioritylasso::prioritylasso() arguments.
  The glmnet::cv.glmnet() pass-through hyperparameters were removed because they were not consistently forwarded (#594).
• surv.cv_glmnet: removed the parameters standardize.response, type.gaussian, type.logistic, and type.multinomial, which are not applicable to the Cox family (#594).
• surv.glmnet: removed the parameters alignment, parallel, type.logistic, and type.multinomial, which are CV-only or not applicable to the Cox family, and renamed the predict parameter predict.gamma to gamma (#594).

Other

• Updated minimum versions of suggested packages, most notably glmnet (>= 5.0) (#594).
• classif.fastai now pins its fastcore Python dependency to a version below 2.0.0 because fastai (<= 2.8.7) is incompatible with fastcore 2.0 but does not declare an upper bound.
• The h2o learners no longer emit a spurious warning about an outdated H2O cluster version when training or predicting.
• surv.cv_glmnet and surv.glmnet: updated for glmnet v5.0, added the train parameters cox.ties (initialized to "breslow" to keep the previous tie-handling behavior), maxp, and path, added the predict parameter exact (surv.cv_glmnet only), fixed predictions for relaxed fits (relax = TRUE), and added a read-only native_model field that returns the fitted glmnet model (#594).
• surv.priority_lasso: added the train parameter cox.ties, initialized to "breslow" (#594).
• surv.survdnn: added the .threads train parameter (#594).
• regr.botorch_fullybayesian now declares its numpyro, jax, and jaxlib Python dependencies so they are installed automatically.
• regr.bart's hyperparameter sigdf was changed to type double.
• regr.crs: added the train parameter max.eval for compatibility with crs (>= 0.15-45) (#601).

New Features

• classif.priority_lasso, regr.priority_lasso, and surv.priority_lasso now support automatic block-priority derivation via adaptive.order = TRUE, following Herrmann et al. (2021), and prediction for automatic block ordering was fixed upstream in prioritylasso version 0.4.0.