OVA Multiclass Classification can be instantiated for variety of sub-trainer training tasks

[rogancarr]

In writing a test for #2859, I have found that it is possible to create an OVA (aka "One-Versus-All") multiclass classification trainer learners other than a binary classifier as an input.

One of our V1 Goals is:

• Metacomponents smartly restrict their use to compatible components. Example: "When specifying what trainer OVA should use, a user will be able to specify any binary classifier. If they specify a regression or multi-class classifier ideally that should be a compile error."

For all types

• Anomaly Detection throws on Fit() due to data mismatch. Suggests that this could work under ideal conditions.
• Binary classification works, as expected.
• Clustering is a runtime error upon pipeline construction (OVA checks the model type produced).
• Multiclass classification is a runtime error upon pipeline construction (OVA checks the model type produced).
• Ranking pipeline can be instantiated, but Fit() fails on "mismatch for label column".
• Regression pipeline can be instantiated, but Fit() fails on "mismatch for label column".

(Note that I didn't run a recommender check because this needs to be coded x64-only at this point.)

Updated: With the recent change to one-label-type-per task, Ranking and Regression now fail on the Fit(). I've updated the title to reflect that we can no longer train working models, but we can instantiate the pipeline.

[ganik]

@rogancarr can you pls share your tests code you used for repro?

[rogancarr]

@ganik Check out this PR for the test cases: #2921