Fix incorrect minimum value in quantile generation

[RAMitchell]

There was a bug for both hist and gpu_hist where the lowest quantile could have an upper bound of zero and lower bound of zero. This fixes it so that the minimum value is strictly less than all other cut values.

[codecov-io]

Codecov Report

Merging #3167 into master will not change coverage.
The diff coverage is 0%.

@@            Coverage Diff            @@
##             master    #3167   +/-   ##
=========================================
  Coverage     45.29%   45.29%           
  Complexity      228      228           
=========================================
  Files           166      166           
  Lines         12901    12901           
  Branches        466      466           
=========================================
  Hits           5844     5844           
  Misses         6865     6865           
  Partials        192      192

Impacted Files	Coverage Δ	Complexity Δ
src/common/hist_util.cc	0.3% <0%> (ø)	0 <0> (ø)	⬇️

---

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 706be4e...23431fa. Read the comment docs.

[hcho3]

Nice! Did you find an actual example where this caused a problem?

[RAMitchell]

It can be reproduced by running the script at #3156.

Specifically it occurred in a column that only contained the values 0 and nan.

[hcho3]

I see. The old behavior doesn't work when the minimum value is exactly zero. I'll approve this PR.