Release: 0.5.2

• Added named pipelines (the new pipelines module and Pipeline class).
  pipeline = classification in the "general" section (or the project's
  section) fills in the settings of the analyst, critic, and artist, so a
  study can be a label, a seed, and a pipeline. The pipelines are
  classification, regression, and count_regression, which compare
  models in an experiment, and classification_inference,
  regression_inference, count_inference, panel_classification, and
  panel_regression, which fit one model for its coefficients (without
  scaling or resampling, and with each category encoded against a
  reference level). Settings that you write win over the pipeline's, a
  worker's own section can name a pipeline for that worker, the pipeline
  sets "task" if it is not set, and the filled-in settings are in
  Project.idea and the exported "settings.json". Add a Pipeline to
  pipelines.PIPELINES to name your own.
• Added three choosers, which the pipelines use and "auto" now resolves to
  in steps named "impute", "encode", and "evaluate" (and in a worker named
  "critic"): auto_imputer fills missing values (with iterative_impute
  when many numeric values are missing, median_impute otherwise, and
  mode_impute for categories); auto_encoder encodes each categorical
  feature by its number of levels (one_hot for 15 or fewer, target for
  more with a label of two classes or a regressed one, count for a label
  of more classes, tfidf for free text, and date_parts for dates); and
  auto_evaluator applies the evaluators that suit the task and the model.
• auto_model knows the needs of ols, logit_sm, probit, and
  mnlogit (such as the number of classes), which it considers when they
  are named in "choices".
• Fixed smote_tomek and smote_enn marking every training row as
  synthetic, because they remove some of the original rows after adding
  rows. Validators then had no real rows to cross-validate, and choosers
  profiled none. Each row they keep that equals an original row is now
  real.
• Fixed fixest predicting missing values (and so a scorecard failing
  with "Input contains NaN") for rows whose level of a fixed effect has no
  estimate: a level that was not in the training rows, or one that
  pyfixest dropped (a singleton, or a level separated by having rows of
  only one class). Such a row is now predicted with that fixed effect at
  the average of its estimated levels. The estimates of the levels are in
  the model's effect_levels_. A fixest model can therefore be
  validated by group_k_fold (each held-out group is predicted with the
  average fixed effect), which used to fail with "could not predict".
• Fixed permutation_importance, shap_partial_dependence, and
  decision_regions failing for a model that needs columns besides its
  features, such as the fixed effects and cluster of a fixest model.
  permutation_importance and shap_partial_dependence hold those columns
  at each row's own values (as shap_importance does), and
  decision_regions holds them at their most common values in the real
  training rows (as it holds the other features at their medians).
  permutation_importance also works for models that are not scikit-learn
  estimators, which scikit-learn's scorers rejected.
• Fixed the report of a project saying that an experiment that conserves
  memory chose "the best of 1 paths". It now gives the number of paths
  compared.
• Inferers (such as dr_learner and partially_linear) now leave out the
  rows that a sampler (such as random_over or smote) added to the
  training rows, which counted some rows more than once and biased the
  estimates and their standard errors. The data is put back afterwards.