Add average precision

[ablaom]

Partly addresses #68.

This PR is ready for review after:

• Add precision-recall curves #67 is merged

as that PR is the base of this one.

Note to self: Rebased off b56bbbf

Docstring

  AveragePrecision()

  Return a callable measure for computing the average precision. Aliases:
  average_precision.

  AveragePrecision()(ŷ, y)

  Evaluate AveragePrecision() on predictions ŷ, given ground truth observations y. It is
  expected that ŷ be a vector of distributions over the binary set of unique elements of
  y; specifically, ŷ should have eltype <:UnivariateFinite from the
  CategoricalDistributions.jl package.

  Average precision is the area under the empirical precision-recall curve, understood as
  a step function. This is to be contrasted with other measures going under the name "area
  under the precision-recall curve", in which the step function is usually replaced by a
  piece-wise linear approximation. Generally, differences between the two are only obvious
  when the number of observations is small, but it is faster to compute average precision.

  Reference: Wikipedia entry, Average precision
  (https://en.wikipedia.org/w/index.php?title=Information_retrieval&oldid=793358396#Average_precision)

  Definition
  ≡≡≡≡≡≡≡≡≡≡

  Adopting each distinct predicted probability p_1, p_2, \ldots, p_k for the positive
  class as a soft probability threshold for predicting an actual class, and assuming these
  thresholds are arranged in decreasing order, we obtain corresponding recalls R_1, R_2,
  \ldots, R_k (monotonically increasing) and precisions P_1, P_2, \ldots, P_k. Adding an
  extra recall, R_{k+1} = 1, the average precision implemented here is defined as

  \sum_{j=1}^k P_j (R_{j+1} - R_j)

  In some other implementations, such as scikit-learn
  (https://scikit-learn.org/stable/modules/generated/sklearn.metrics.average_precision_score.html#sklearn.metrics.average_precision_score),
  P_j is replaced by P_{j+1}. However, this requires the definition of a precision for
  unit recall, in the case the predicted positive class probabilities exclude 1.0, and
  this is avoided here.

  Core implementation: Functions.average_precision.

  Generally, an observation obs in MLUtils.eachobs(y) is expected to satisfy
  ScientificTypes.scitype(obs)<:ScientificTypesBase.OrderedFactor{2}.

  See also precision_recall_curve.

  For a complete dictionary of available measures, keyed on constructor, run measures().

  Traits
  ≡≡≡≡≡≡

  consumes_multiple_observations = true
  can_report_unaggregated = false
  kind_of_proxy = LearnAPI.Distribution()
  observation_scitype = ScientificTypesBase.OrderedFactor{2}
  can_consume_tables = false
  supports_weights = false
  supports_class_weights = false
  orientation = StatisticalMeasuresBase.Score()
  external_aggregation_mode = StatisticalMeasuresBase.Mean()
  human_name = average precision

[ablaom]

cc @DilumAluthge

[ablaom]

Closed in favour of #72.