Fix KMeans seeding consistency across n_jobs (swev-id: scikit-learn__scikit-learn-9288)

[casey-brooks]

Summary

• ensure KMeans precomputes per-run seeds so multi-worker runs consume identical RNG draws
• add regression coverage to compare inertia across n_jobs for both estimator and function APIs

Testing

• source /workspace/.venv/bin/activate
• export PATH=$HOME/.nix-profile/bin:$PATH
• export LD_LIBRARY_PATH=$(dirname $(g++ -print-file-name=libstdc++.so.6)):$HOME/.nix-profile/lib
• pytest sklearn/cluster/tests/test_k_means.py

[casey-brooks]

Local Testing

• source /workspace/.venv/bin/activate
• export PATH=$HOME/.nix-profile/bin:$PATH
• export LD_LIBRARY_PATH=$(dirname $(g++ -print-file-name=libstdc++.so.6)):$HOME/.nix-profile/lib
• pytest sklearn/cluster/tests/test_k_means.py # 127 passed, 0 failed, 0 skipped (12573 warnings)

[noa-lucent]

Precomputing the per-run seeds keeps the sequential and parallel paths in sync, and the new regression tests exercise both estimator and functional APIs. LGTM.