[tracking] PyCall -> PythonCall transition

[ablaom]

In discussions with @tylerjthomas, @OkonSamuel and others, we have a tentative plan to transition model-providing packages that use PyCall to ones that use PythonCall instead. Unfortunately, loading packages that use both simultaneously, while possible, requires interventions beyond the average user (in my view) and complicates integration tests and updating the MLJ Model Registry. So this needs to be a coordinated transition.

A tentative plan is:

• [x] finalise a PythonCall branch for MLJScikitLearnInterface.jl (which also drops ScikitLearn.jl as dependency): https://github.com/JuliaAI/MLJScikitLearnInterface.jl/pull/56
• [x] finalise a PythonCall branch for OutlierDetectionPython: https://github.com/OutlierDetectionJL/OutlierDetectionPython.jl/pull/3
• [x] check for any other PyCall dependencies in packages registered in the MLJ Model Registry (@ablaom)
• [x] run tests from MLJTestIntegration, locally, based on the PythonCall branches
• [x] take several deep breaths (@ablaom)
• [x] merge the MLJScikitLearnInterface.jl branch into master and tag new release
• [x] merge the OutlierDetectionPython.jl branch into master and tag new release
• [x] update MLJ Model Registry, including CatBoost.jl and tag a new MLJModels.jl release (@ablaom)

[tylerjthomas9]

• [ ] take several deep breaths (@ablaom)

The most important step

[ablaom]

Okay. I'm taking those deep breaths now. Will roll the rest of this out (including CatBoost.jl addition to registry) after the weekend.