Clip small positive values in gamma-nloglik

[gorkemozkaya]

For the gamma-nloglik eval metric, small positive values in the labels are causing NaN's in the outputs, as reported here: #5349. This will add clipping on them, similar to what is done in other metrics like poisson-nloglik and logloss.

[codecov-io]

Codecov Report

Merging #6537 (0ed3370) into master (fbb980d) will increase coverage by 80.18%.
The diff coverage is n/a.

@@             Coverage Diff             @@
##           master    #6537       +/-   ##
===========================================
+ Coverage        0   80.18%   +80.18%     
===========================================
  Files           0       13       +13     
  Lines           0     3558     +3558     
===========================================
+ Hits            0     2853     +2853     
- Misses          0      705      +705     

Impacted Files	Coverage Δ
python-package/xgboost/__init__.py	89.47% <0.00%> (ø)
python-package/xgboost/data.py	57.05% <0.00%> (ø)
python-package/xgboost/compat.py	72.54% <0.00%> (ø)
python-package/xgboost/libpath.py	55.55% <0.00%> (ø)
python-package/xgboost/tracker.py	93.98% <0.00%> (ø)
python-package/xgboost/callback.py	80.33% <0.00%> (ø)
python-package/xgboost/config.py	100.00% <0.00%> (ø)
python-package/xgboost/training.py	95.65% <0.00%> (ø)
python-package/xgboost/rabit.py	76.19% <0.00%> (ø)
python-package/xgboost/sklearn.py	89.48% <0.00%> (ø)
... and 3 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update fbb980d...0ed3370. Read the comment docs.

[trivialfis]

Thanks for the PR!