Create a class hierarchy for proxy-loss types: Quadratic/XEntropy/Hinge

[AndreFCruz]

Summary

• Currently we have to do string comparisons multiple times per data row, per iteration, to figure out which proxy loss function to use to compute the gradients.
  • e.g., https://github.com/feedzai/fairgbm/blob/bulk-fairgbm-modifications/src/objective/constrained_recall_objective.hpp#L86
• According to our benchmarks, this can take up around 30% of the total training time.
• We should create a common interface w/ different subclasses to optimize this.
• This has already been started in FairGBM refactor (multi-threading etc.) #18

[AndreFCruz]

The skeleton for this refactor has been started: I've created a common interface for all proxy losses at include/LightGBM/utils/loss_proxies.hpp

[AndreFCruz]

Closed by #35