Experimental sparse/MNMG logistic regression

[JohnZed]

For 0.12, this can still be experimental. But we want to demonstrate supporting:

• Dask-distributed data
• MNMG logistic
• A minibatch style solver (e.g. ADAM)
• Sparse, one-hot-encoded data

Optionally we can add support for:

• Easy integration of embeddings for categorical features

[daxiongshu]

Done:

• fixed slowdown issues due to fit_intercept and normalize when fitting dask_glm.
• end-to-end dask-cudf ETL + dask_glm runs without errors on criteo 1tb benchmark.

To do:

• check correctness of dask_glm against glm models of dask_ml and sklearn.
• measure the speedup of dask_glm on criteo data over cpu baseline.

[JohnZed]

Initial commits to dask-glm:

• [Review] Allow lbfgs and admm with dask cupy inputs dask/dask-glm#89
• Fix algorithms "newton", "gradient_descen" and "proximal_grad" with cupy inputs dask/dask-glm#87

[beckernick]

I believe #3512 closed this issue. Is that correct @daxiongshu ?

[daxiongshu]

yes, thank you!