Mlrl-boomer

A major update to the BOOMER algorithm that features the following changes:

- Large parts of the code (loss functions, calculation of gradients/Hessians, calculation of predictions/quality scores) have been refactored and rewritten in C++. This comes with a constant speed-up of training times.
- Multi-threading can be used to parallelize the evaluation of a rule's potential refinements across multiple CPU cores.
- Sparse ground truth label matrices can now be used for training, which may reduce the memory footprint in case of large data sets.
- Additional parameters (`feature-format` and `label-format`) that allow to specify the preferred format of the feature and label matrix have been added.

A major update to the BOOMER algorithm that features the following changes:

- Includes many refactorings and quality of live improvements. Code that is not directly related with the algorithm, such as the implementation of baselines, has been removed.
- The algorithm is now able to natively handle nominal features without the need for pre-processing techniques such as one-hot encoding.
- Sparse feature matrices can now be used for training and prediction, which reduces the memory footprint and results in a significant speed-up of training times on some data sets.
- Additional hyperparameters (`min_coverage`, `max_conditions` and `max_head_refinements`) that provide fine-grained control over the specificity/generality of rules have been added.

The first version of the BOOMER algorithm used in the following publication:

*Michael Rapp, Eneldo Loza Mencía, Johannes Fürnkranz and Eyke Hüllermeier. Gradient-based Label Binning in Multi-label Classification. In: Proceedings of the European Conference on Machine Learning and Knowledge Discovery in Databases (ECML-PKDD), 2021, Springer.*

This version supports the following features to learn an ensemble of boosted classification rules:

- Different label-wise or example-wise loss functions can be minimized during training (optionally using L2 regularization).
- The rules may predict for a single label, or for all labels (which enables to model local label dependencies).
- When learning a new rule, random samples of the training examples, features or labels may be used, including different techniques such as sampling with or without replacement.
- The impact of individual rules on the ensemble can be controlled using shrinkage.
- The conditions of a recently induced rule can be pruned based on a hold-out set.
- The algorithm currently only supports numerical or ordinal features. Nominal features can be handled by using one-hot encoding.