Use double-precision to store split thresholds

[hcho3]

One of the model builder tests failed because scikit-learn uses double-precision floating-point to store split points, whereas Treelite was using single-precision. This proved to be a problem because the threshold 4.7999999523 was changed to 4.8000001907 when it was cast to single-precision.

Fix: Use double-precision to store split thresholds.

[codecov-io]

Codecov Report

Merging #55 into master will increase coverage by 0.42%.
The diff coverage is 100%.

@@             Coverage Diff              @@
##             master      #55      +/-   ##
============================================
+ Coverage     87.31%   87.73%   +0.42%     
  Complexity       35       35              
============================================
  Files            76       76              
  Lines          5225     5233       +8     
  Branches         21       21              
============================================
+ Hits           4562     4591      +29     
+ Misses          642      621      -21     
  Partials         21       21

Impacted Files	Coverage Δ	Complexity Δ
python/treelite/frontend.py	74.9% <ø> (ø)	0 <0> (ø)	⬇️
src/c_api/c_api.cc	85.14% <ø> (ø)	0 <0> (ø)	⬇️
include/treelite/base.h	50% <ø> (ø)	0 <0> (ø)	⬇️
tests/python/test_model_builder.py	100% <100%> (ø)	0 <0> (ø)	⬇️
src/frontend/protobuf.cc	90.74% <100%> (+9.72%)	0 <0> (ø)	⬇️
include/treelite/common.h	83.2% <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 c9023e7...b7972fb. Read the comment docs.