Reorganization of the predictor types

[DilumAluthge]

Currently, we only have a single predictor type:

struct SossMLJPredictor{M} <: Distributions.Distribution{MixedVariate, MixedSupport}
    model::M
    post
    pred
    args
end

I think we'll need a slightly more complicated set of types.

For predict_joint, we will continue to use SossMLJPredictor. But for predict, I think we'll need different types.

For predict with a classification model, we'll need to use MLJBase.UnivariateFinite for compliance with the MLJ model interface.

For predict with a regression model, it's not immediately obvious to me what we should return.

[cscherrer]

Yeah, I think you're right. How about this:

• We keep SossMLJPredictor as the "general" case, which is probably also good for regression
• Add SossMLJMarginalPredictor for marginals. I think users shouldn't have to type this, we can have a marginalize function to build it
• MLJSossClassifier for classification in the MLJ sense, since that needs this special treatment

Oh, and we'll need type parameters on these, I just left it open to get a quick start

[DilumAluthge]

For consistency, can we name the last one SossMLJClassifier?

[cscherrer]

Oops, yes that's what I meant :)

[DilumAluthge]

The parts of this that are blocking 0.1.0 are implemented in #91.

There is still additional work to be done, but it doesn't block 0.1.0.

[cscherrer]

I'm now second-guessing whether we really need a SossMLJClassifier type. Let's see how it goes first

[DilumAluthge]

I'm going to mark this as potentially breaking, since it will probably result in some changes to the return types of public functions.