Stochastic Gradient Trees

[henrygouk]

I have adapted/mostly reimplemented the Stochastic Gradient Tree method from my ACML 2019 paper. This implementation contains a couple of minor enhancements:

• Leaves can have vector valued predictions, so there is no longer a need to do one-vs-rest (or similar) transforms for multiclass problems.
• I implemented the resplit strategy from EFDT. Previously, SGT would do the same approach as EFDT for choosing a split, but it would not reconsider splits as more data became available.
• I removed the gamma regulariser. This did not seem to have much of an impact in practice.
• This implementation is capable of classification, regression, multi-label classification, and multi-target regression without using any problem transformation strategies.