RF: Add Gamma and Inverse Gaussian loss criteria (rapidsai#4216)

This PR adds the Gamma and Inverse Gaussian Criteria to train decision trees, along with modifications to rf unit tests.


---


checklist:
- [x] Add Gamma and Inverse Gaussian Objective classes
- [x] Add C++ tests for above
- [x] Add remaining C++ tests for other objective functions: entropy and mean squared error
- [x] Add python level convergence tests for gamma and inverse gaussian ( just like the one added for poison loss in rapidsai#4156 )
- [x] Check for regressions by benchmarking on gbm-bench
- [x] Convergence plots showing model trained on particular criteria performs better on it's own loss metric than a baseline (`mse`)

Authors:
  - Venkat (https://github.com/venkywonka)

Approvers:
  - Rory Mitchell (https://github.com/RAMitchell)
  - William Hicks (https://github.com/wphicks)
  - Dante Gama Dessavre (https://github.com/dantegd)

URL: rapidsai#4216

cpp/include/cuml/tree/algo_helper.h

@@ -23,6 +23,8 @@ enum CRITERION {
   MSE,
   MAE,
   POISSON,
+  GAMMA,
+  INVERSE_GAUSSIAN,
   CRITERION_END,
 };