Release: 0.4.1

• regression_table takes "sets" of features (such as the dummy variables
  of one category, as "family, partner, stranger; under_12, teen" in
  settings), whose members' average marginal effects are each the change
  from the set's reference (every member 0) to that member alone, rather
  than from moving each dummy on its own, which made rows that cannot
  happen.
• Added techniques that wrap mlxtend, which the new mlxtend extra (and
  all) installs: sequential_selection and exhaustive_selection,
  reducers that search for the set of features with which a model (a
  logistic or linear regression, or "estimator") scores best in
  cross-validation, folding by the dataset's first group as a model's search
  does, and store the score of each set as "{name}_search";
  association_rules, a describer of which values (of the categorical and
  boolean columns, and of a classified label) tend to appear in the same
  rows; bias_variance, an evaluator that splits the model's expected loss
  on the test rows into its bias and variance; and decision_regions, a plot
  of the class that a classifier predicts across two features.
• Added dtreeviz's plots of decision trees, which the explain extra now
  installs: tree_leaf_sizes, tree_leaf_distributions, tree_leaf_purity,
  tree_feature_space, and tree_prediction_path. They draw a
  decision_tree or one tree of a random_forest, extra_trees, or
  adaboost model (and tree_leaf_sizes and tree_leaf_distributions one
  tree of an xgboost or lightgbm model), in the style of the other
  figures. dtreeviz's drawing of a whole tree is not wrapped, since it needs
  Graphviz and makes only SVG files; tree_plot draws one.
• models.search_folds returns the folds of a hyperparameter search (which
  Model._folds now calls), so that other techniques can fold by groups as
  a search does.
• Plots that draw with matplotlib.pyplot (shap's, and now mlxtend's and
  dtreeviz's) are lent their figure by plots._lent, which also keeps
  pyplot.show from opening a window while they draw.
• A model's hyperparameter search ("grid", "random", or "optuna") keeps the
  rows of each group of the dataset's first group column in one fold, as the
  group validators do, when the dataset has groups: with
  StratifiedGroupKFold for a classifier and GroupKFold for a regression.
  Before, its folds divided the rows at random, so rows of one group (such as
  the votes of the judges of one case) could be on both sides of a fold and
  the search scored a model on rows it had partly seen. A row with a missing
  group is a group of its own.
• Brought in the CI fixes of snickerdoodle 0.4.3: GitHub releases are made
  with the GitHub CLI (in place of the deprecated gh-action-auto-release
  Action), with the released version's section of CHANGELOG.md as their
  notes, only from main and after the tests and linters pass. The test
  matrix now runs every Python version it names, instead of recreating the
  environment with Python 3.11 at uv run.