Enhancement of BinaryThresholdPredictor to handle outlier detection models

[ablaom]

This PR continues #399 and:

• Expands the types of probabilistic detectors that can be wrapped by BinaryThresholdPredictor, so as to include new outlier detector types added to MLJModelInterface here.
• (breaking) Changes the field name wrapped_model to model, for consistency with the wrappers TunedModel and IteratedModel. I haven't kept backwards compatibility as a I don't think this is going to effect many users yet, but I am tagging the release as minor (breaking). @OkonSamuel may want to comment.
• Removes a check and test that are redundant. From the original test code:

function MMI.predict(model::BinaryThresholdPredictor, fitresult, X)
   yhat = MMI.predict(model.wrapped_model, fitresult[1], X)
   length(yhat.prob_given_ref) == 2 || begin
   # Due to resampling it's possible for Predicted `AbstractVector{<:UnivariateFinite}`
   # to contain one class. Hence the need for the following warning
   @warn "Predicted `AbstractVector{<:UnivariateFinite}`"*
       " contains only 1 class. Hence predictions will only contain this class "*
       "irrrespective of the set `threshold` "

Actually, it is impossible for a correctly implemented probalistic classifier to predict UnivariateFinite having a different number of classes from the target on which it was trained, because the pool of the UnivariateFinite should always coincide with the pool of the target on which the model was trained and resampling does not effect the pool of the training target, as in this example:

julia> y = categorical([1, 1, 1, 2]);

julia> y[1:2]
2-element CategoricalArray{Int64,1,UInt32}:
 1
 1

julia> levels(y[1:2])
2-element Vector{Int64}:
 1
 2

cc @davnn

@OkonSamuel Be great if you can have a look over this. Do you when you might get a chance, assuming your'e willing?

[codecov-commenter]

Codecov Report

Merging #400 (4adc491) into dev (2f67f4b) will decrease coverage by 0.53%.
The diff coverage is 94.23%.

@@            Coverage Diff             @@
##              dev     #400      +/-   ##
==========================================
- Coverage   75.56%   75.02%   -0.54%     
==========================================
  Files          12       12              
  Lines         937      981      +44     
==========================================
+ Hits          708      736      +28     
- Misses        229      245      +16     

Impacted Files	Coverage Δ
src/builtins/ThresholdPredictors.jl	95.45% <94.23%> (-4.55%)	⬇️
src/metadata.jl	79.41% <0.00%> (-2.31%)	⬇️
src/builtins/Transformers.jl	96.37% <0.00%> (-0.17%)	⬇️
src/utilities.jl	0.00% <0.00%> (ø)
src/registry/src/Registry.jl	0.00% <0.00%> (ø)
src/registry/src/check_registry.jl	0.00% <0.00%> (ø)
src/loading.jl	62.35% <0.00%> (+0.44%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 2f67f4b...4adc491. Read the comment docs.

[ablaom]

@davnn @OkonSamuel

Wanting to move forward with this, I have added backwards compatibility to BinaryThresholdPredictor (with a depwarn for attempts to access using wrapped_model instead of model) and down-graded the new version number to a patch.