Add Hard Voting and Median Ensembler as well as new Task Evaluator

[Heptasphere]

First PR for #334.

Description of changes

Two new ensemblers and a task evaluator were added:

• MedianEnsembler: Ensembler that uses the median rule (robust soft voting in
• HardVotingEnsembler: Ensembler for classification only that uses hard
• NoPreprocessingTaskEvaluator: Task Evaluator allowing ensemblers that use all predicted classes in multiclass training/prediction, even for fast metrics, not just extracted true classes.

NoPreprocessingTaskEvaluator doesn't preprocess the base learner predictions before training / prediction, it only preprocesses the ensembler results for error estimation.

Testing

All tests are in tests/tabarena/simulation/test_ensembler.py

Basic tests for the format of the ensemblers were added. Tests for their results on a toy example were also added.

Test to check that the new task evaluator had similar results to TaskEvaluator on a simple ensembler (TopKAverageEnsembler) was added.

• In the case of multiclass, small changes to the metric value were expected due to the particularities of log_loss preprocessing. We use the original float32 rather than the preprocessed float64, leading to a $2^{-23} \approx 10^{-6}$ estimation difference for the metric.
• This task evaluator is not expected to work well with Ensemble Selection.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[Heptasphere]

The code for the MedianEnsembler could easily have been made into a general code for min/max/median/whatever rule through giving a function as argument and normalizing for binary classification (unnecessary for median due to its symmetry).

I considered that median is already not very popular so any other rule might remain forever unused. (That, and I'm always a bit paranoid about letting a method defined by the user be an input for anything.)