[Tracking issue] Add raw_training_scoresaccessor function

[ablaom]

To address #841 (comment) I suggest we add a raw_training_scores accessor function, following the established pattern for feature_importances and training_losses.

@davnn Are you happy with the name of this function? I'm thinking here of consistency with training_losses.

Here are the steps (adapted from #747).

Step 1. Prepare a bunch of PR's to be rolled out after testing, all called scores:

• Add reports_raw_training_scores trait to StatisticalTraits, defaulting to false, as here for feature_importances (but need to also add to list of all traits at top of file)
• Add raw_training_scores(model, fitresult, report) stub to MLJModelInterface (in model_api.jl), as here; fallback to return nothing.
• Overload MLJModelInterface.raw_training_scores(mach::Machine) following this pattern and re-export raw_training_scores (the reports_raw_training_scores trait will be automatically re-exported in MLJBase and MLJ - no action required)
• In MLJ: re-export raw_training_scores

Step 2

• @blaom to check MLJTestIntegration runs with no new surprises (and maybe adds some detector-specific testing)

Step 3

• Merge and rollout the scores PRs and tags new releases, taking care to bump compats accordingly. Probably @ablaom to do.
• Update MLJ model API docs

Step 4

• Roll out implementations for packages that already report
  scores (in, eg, their report); @davnn to complete this checklist:
  • OutlierDetection.jl
  • ?