Add ordered_complement helper, vectorize diff branch, and tweak vignette headers

[OVVO-Financial]

Motivation

• Avoid the cost and ordering effects of np.setdiff1d in resampling code while preserving the original index order for training/validation splits.
• Improve performance and clarity of the vector-branch reduction in _dy_d_best_plan by removing looped index bookkeeping and using array reshaping and vectorized operations.
• Clean up vignette headers to better reflect the source R markdown filenames.

Description

• Add src/nns/_indices.py implementing ordered_complement(all_index, excluded) which returns the values of all_index not present in excluded while preserving input order.
• Replace usages of np.setdiff1d(all_index, ...) with ordered_complement in src/nns/boost.py and src/nns/stack.py for training/validation index construction.
• Refactor the vector branch in src/nns/diff.py::_dy_d_best_plan to compute per-group means by reshaping the prediction block and compute First and Second arrays with vectorized arithmetic, removing position/ids bookkeeping and looped element-wise aggregation.
• Update vignette scripts (examples/vignettes/03_correlation_and_dependence.py and 04_normalization_and_rescaling.py) to adjust header comments and docstrings and tidy imports.
• Add unit test tests/invariants/test_indices.py asserting that ordered_complement preserves the source index order.

Testing

• Ran the new unit test pytest tests/invariants/test_indices.py and it passed.

---

Codex Task