[BUG] Divergent KMeans++ Behavior for Equal Length Host and Device Resident Arrays

A device array calls the sequential kmeans++ algorithm with the entire dataset. However a host array with device_buffer_samples = 0 (default) would take a sample of shape min(3 * n_clusters, n_samples) * n_cols. This calls a cuml.accel test to fail.

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
- generated xml file: /__w/cuml/cuml/test-results/junit-cuml-accel-scikit-learn.xml -
=========================== short test summary info ============================
FAILED python/cuml/cuml_accel_tests/upstream/scikit-learn/tests/test_pipeline.py::test_fit_predict_on_pipeline - AssertionError: 
Arrays are not almost equal to 6 decimals

Mismatched elements: 120 / 150 (80%)
First 5 mismatches are at indices:
 [0]: 7 (ACTUAL), 1 (DESIRED)
 [1]: 1 (ACTUAL), 4 (DESIRED)
 [2]: 1 (ACTUAL), 4 (DESIRED)
 [3]: 1 (ACTUAL), 4 (DESIRED)
 [4]: 7 (ACTUAL), 1 (DESIRED)
Max absolute difference among violations: 6
Max relative difference among violations: 6.
 ACTUAL: array([7, 1, 1, 1, 7, 3, 7, 7, 1, 1, 7, 7, 1, 1, 3, 3, 3, 7, 3, 7, 7, 7,
       7, 7, 7, 1, 7, 7, 7, 1, 1, 7, 3, 3, 1, 1, 7, 7, 1, 7, 7, 1, 1, 7,
       7, 1, 7, 1, 7, 7, 6, 6, 6, 5, 6, 0, 6, 5, 6, 0, 5, 0, 5, 6, 0, 6,...
 DESIRED: array([1, 4, 4, 4, 1, 7, 1, 1, 4, 4, 7, 1, 4, 4, 7, 7, 7, 1, 7, 7, 1, 7,
       1, 1, 1, 4, 1, 1, 1, 4, 4, 1, 7, 7, 4, 4, 1, 1, 4, 1, 1, 4, 4, 1,
       7, 4, 7, 4, 7, 1, 6, 6, 6, 2, 6, 2, 6, 2, 6, 2, 2, 6, 2, 6, 2, 6,...
= 1 failed, 37461 passed, 6138 skipped, 662 xfailed, 101 xpassed, 5277 warnings in 1086.09s (0:18:06) =

Start with python/cuml/cuml_accel_tests/upstream/scikit-learn/tests/test_pipeline.py::test_fit_predict_on_pipeline and reproduce the reported failure. Trace the KMeans++ handling for host arrays with device_buffer_samples=0 and device-resident arrays, then verify that equal-length inputs produce consistent results and that the test passes.