Add demo test for multioutput_reduced_gradient.py

[Nityahapani]

Add demo test for multioutput_reduced_gradient.py

Summary

test_demos.py runs most demo scripts as a real subprocess smoke test.
multioutput_reduced_gradient.py had no coverage. This is genuinely the
only thing exercising xgboost.objective.TreeObjective at all (grepped
tests/python/*.py for both the script name and TreeObjective directly —
zero hits either way), so this is the sole safety net for that experimental
feature, not just a routine smoke test.

Change

+@pytest.mark.skipif(**tm.no_sklearn())
+def test_multioutput_reduced_gradient_demo() -> None:
+    # Runs with the default `--device cpu`; `cupy` is only imported on the
+    # (unused here) `--device cuda` code paths.
+    script = os.path.join(PYTHON_DEMO_DIR, "multioutput_reduced_gradient.py")
+    cmd = [PYTHON, script]
+    subprocess.check_call(cmd)

Verification

• Runs with the script's default --device cpu (no args needed)
• Its three import cupy statements are all inside else branches gated on
  self.device == "cpu" checks — confirmed by reading each one's
  surrounding context directly — so none of them execute on this path, no
  cupy/GPU dependency for this test
• no_sklearn() skip marker matches every other sklearn-dependent test in
  this file (the script uses sklearn.datasets.make_regression,
  sklearn.base.BaseEstimator, and sklearn.decomposition.TruncatedSVD,
  all part of core scikit-learn, no extra dependency)

I also checked the other 8 demo scripts currently missing test coverage
(cat_in_the_dat.py, cover_type.py, gamma_regression.py,
gpu_tree_shap.py, learning_to_rank.py, quantile_data_iterator.py,
spark_estimator_examples.py, update_process.py) and confirmed each is
correctly excluded for a real reason (local-only data files, GPU/RAPIDS
dependencies, a large external ranking dataset, or pyspark) — not further
oversights, so this PR is scoped to just the one genuine gap.