[REVIEW] SVM class and sample weights

[tfeher]

This PR adds SVM class and sample weights, and closes issue #2222.

• The Python layer accepts both sample and class weights.
• The C++ layer accepts only sample weights. Class weight can be included by setting sample weights based on the class label (as it is done here).

The weights scale the penalty parameter C for each sample individually. The essence of this PR is the two code lines change in SmoBlockSolver, the rest just makes sure that the correctly weighted penalty parameter reaches that point, and the results are correctly processed afterwards.

Additional unrelated change: the unused fit method from SVMBase was removed (#2409).

[GPUtester]

Please update the changelog in order to start CI tests.

View the gpuCI docs here.

[tfeher]

Thanks @teju85 for the review, I have fixed the issues!

[teju85]

Thank you @tfeher for the PR. Changes LGTM.

[dantegd]

As part of CI the doxygen docs are now checked, the parameter sample_weight seems to be missing from the docstring now:

Generating docs for compound ML::Internals::Cal/jenkins/workspace/rapidsai/gpuci/cuml/prb/cuml-gpu-build/cpp/include/cuml/svm/svc.hpp:49: error: The following parameter of ML::SVM::svcFit(const cumlHandle &handle, math_t *input, int n_rows, int n_cols, math_t *labels, const svmParameter &param, MLCommon::Matrix::KernelParams &kernel_params, svmModel< math_t > &model, const math_t *sample_weight=nullptr) is not documented:

  parameter 'sample_weight' (warning treated as error, aborting now)

[tfeher]

Thanks @dantegd, I have corrected the docstring!