Update export_model(): Add split_kind field

[chyunsu3]

Fixes #696

@betatim Can you review? I tested the change using the reproducer script from #696.

[codecov]

Codecov Report

❌ Patch coverage is 60.00000% with 4 lines in your changes missing coverage. Please review.
✅ Project coverage is 84.35%. Comparing base (3b2a9a3) to head (9307469).

Files with missing lines	Patch %	Lines
python/treelite/sklearn/exporter.py	60.00%	4 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff              @@
##           mainline     #698      +/-   ##
============================================
- Coverage     84.39%   84.35%   -0.04%     
============================================
  Files            75       75              
  Lines          6862     6867       +5     
  Branches        557      557              
============================================
+ Hits           5791     5793       +2     
- Misses         1071     1074       +3     

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[betatim]

Looks good to me.

One thing I'd add is a check that the fields in the imported node dtype match what treelite knows. Because the fields need filling treelite needs to know about them. I think right now it would maybe work but give wrong results silently.

Something like this at the top of exporter.py

# Fields of scikit-learn's private Node struct that this exporter knows how to
# handle. If scikit-learn grows a field outside this set we cannot populate it,
# and since we now allocate with scikit-learn's own NODE_DTYPE the dtype check in
# Tree.__setstate__ will no longer catch it -- the field would silently keep its
# zero value. `left_cat_bitset` is knowingly left zeroed: trees with categorical
# splits are rejected before we allocate.
_KNOWN_NODE_FIELDS = frozenset(
    {
        "left_child",
        "right_child",
        "feature",
        "threshold",
        "left_cat_bitset",
        "impurity",
        "n_node_samples",
        "weighted_n_node_samples",
        "missing_go_to_left",
        "split_kind",
    }
)

and then a check like this in _export_tree:

unknown_fields = sorted(set(NODE_DTYPE.names) - _KNOWN_NODE_FIELDS)
if unknown_fields:
    raise NotImplementedError(
        f"scikit-learn {sklearn_version} stores tree node fields that this "
        f"version of Treelite does not know how to populate: {unknown_fields}. "
        "Exporting would produce a model that predicts incorrectly. "
        "Please upgrade Treelite."
    )