Multi-class support for LogisticRegression

[rikturr]

(splitting out a request from #386 )

LogisticRegression currently only supports binary classification (the multi_class argument is ignored). This feature request is to add multi-class support!

[TomAugspurger]

Thanks.

I think the main issue here is likely around inferring whether a problem is binary or multi-class. I'd prefer to avoid inference at fit time. Users would be explicit about it (so that different partitions don't infer different types).

[yniu87]

From the perspective of science, it's important to have multi-class support instead of nesting several binary logistic regressions. (see discussion here). I'd like to work on adding multinomial logistic regression.

@TomAugspurger

1. It seems like I need to first work on https://github.com/dask/dask-glm/tree/main/dask_glm. Adding new distributions under families.py, new loss functions under algorithms.py, etc...
2. After step 1 is done, I can start to modify glm.py inside dask-ML. Does this process sound correct to you? Are you open to accepting these PRs in the future?

Thank you!

[khyll]

Seconding wish for multiclass logistic regression. Keep up the great work!