Relax restrictions on model type in resampling (evaluate!)

[ablaom]

Current behaviour

Currently a model in evaluate/evaluate! must subtype Supervised or Annotator. However, in the current implementation, all that's essential for the model is:

• It implements a predict method
• It's fit method consumes a target y (which must be provided in the evaluate call, or machine constructor).
• In the common case that operation is left unspecified, MLJModelnterface.prediction_type must be one of: :deterministic, :probabilistic, :interval.

Subtypes of Annotator and Supervised (e.g,Probabilistic, Deterministic, Interval) have the obvious prediction_type fallbacks, but the fallback for Unsupervised models is :unknown.

What this PR does

This PR:

• Removes the current model type restriction: Now any <:Model is now allowed.
• Adds the currently missing check that a target y has been provided.
• In the case that an operation is not specified, and the prediction_type is not one of :deterministic, :probabilistic, :interval, the error which is already thrown in that case is made more informative. It now suggests two remedies: (i) explicitly specify operation(s), or (ii) post an issue to have the model's prediction_type reviewed.

Context

More and more we are relying on traits and less on the abstract model hierarchy. Currently, however, only Unsupervised models can have their transform method propagated in a Pipeline (and predict is propagated by Supervised models). So some models want to be Unsupervised for pipeline use, despite the fact they are also "supervised" - a case in point is RecursiveFeatureElimination (currently Supervised and so not pipelinable).

To do

Just to be sure, let's wait for

• Run local MLJ integration tests (@ablaom)

Afterword

We should ultimately like to evaluate models that do not consume a target in training, such as clustering models. In that case, the target is only supplied for the test sets (e.g., for computing the Rand index). However, that requires a more substantial redesign of evaluate.

[OkonSamuel]

Looks Good to me!!!.