[FEA] SVM working set selection using priority

[tfeher]

Is your feature request related to a problem? Please describe.
The working set selection algorithm fundamentally influences the number of outer iterations in the SMO algorithm and therefore the total training time.

Currently we use a FIFO strategy similar to ThunderSVM. An alternative priority based selection is also implemented experimentally, but it is not possible to use that currently.

Describe the solution you'd like

• Test the priority based working set selection method.
• Make it possible to choose between the available working set selection algorithms.
• Compare the performance and select the fastest as default.

Describe alternatives you've considered
The current default FIFO strategy works reasonably well.

Additional context
https://github.com/tfeher/cuml/blob/1d64dbe37835e106e22b3179071618204b517b18/cpp/src/svm/workingset.h#L227