Add ability to export as sklearn RF

[hcho3]

This PR adds a function to export Treelite models as scikit-learn random forests.
It makes use of the field accessor API.

Usage:

# tl_model is an treelite.Model object

# Returns either RandomForestClassifier / RandomForestRegressor
clf = treelite.sklearn.export_model(tl_model)  

Limitations

• Export to GradientBoosting is not yet supported.
• Treelite models with categorical splits are not yet supported.
• Test coverage currently ensures that predict() and predict_proba() return correct result. Other estimator metadata may be missing or incorrect.
• Most of export logic is written in the Python layer so may be slow for very large models.

[codecov]

Codecov Report

Attention: Patch coverage is 90.30303% with 16 lines in your changes missing coverage. Please review.

Project coverage is 84.87%. Comparing base (386f50c) to head (36cf0bf).
Report is 1 commits behind head on mainline.

Files with missing lines	Patch %	Lines
python/treelite/sklearn/exporter.py	85.58%	16 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff              @@
##           mainline     #587      +/-   ##
============================================
+ Coverage     84.80%   84.87%   +0.06%     
============================================
  Files            75       77       +2     
  Lines          6587     6743     +156     
  Branches        528      531       +3     
============================================
+ Hits           5586     5723     +137     
- Misses         1001     1020      +19     

Flag	Coverage Δ
	?

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[wphicks]

Looks great! One question for my own benefit, but it should not block merge.

[betatim]

Looks reasonable to me. Left a few comments which are mostly for my education.

It could be good to add some tests to trigger the various "this model can't be exported to scikit-learn" conditions. That way it is tested and we make codecov happy.

Could also be worth having one test that uses a round trip from scikit-learn to scikit-learn to check that all the fitted attributes (the ones ending in _) are set to the right value (or not set if not supported)

[hcho3]

Created #588 to remind myself to remove the pin cpplint==1.6.1 later.