Add short-series sampling and band scripts.

[rabujamra]

Adds the two reproducibility scripts from #114: the original short-series sampling study and the n=3–30 band extension. No library behavior or methodology is changed.

Summary

• What: Adds validation/short_series_sampling.py and validation/short_series_bands.py.
• Why: Preserves the reproducibility work from Series length (T) as an axis in design-state admissibility — and a saturation question #114 in the repository and supplies the extended n=3–30 sampling table requested for the short-series methodology discussion.
• Scope: Validation/reproducibility only. No package code, public API, estimator, design-state logic, or chart-selection behavior is changed.

Contract / invariants

Check the ones this change had to keep true. Anything you cannot check, explain below.

• Defaults unchanged — existing calls produce equivalent results
• Residuals unaffected — no residual or design-state recomputation
• Row/index alignment preserved — output row count and ordering unchanged
• Output schema compatible — column names/types and the stats-dict shape
  ({N, center, lpl, upl}) still hold
• Pinned error strings unchanged — or the tests that pin them are updated
  deliberately, and the commit says why

These invariants are not exercised by this PR because it adds standalone
validation/reproducibility scripts only and does not modify package code.

Behaviour changes

None. Validation/reproducibility scripts only.

Methodology

• No methodology change — implementation, ergonomics, or docs only
• Methodology change, validated against Bishop's reference data
  (cite the chapter/equation/table)

Tests

• pytest tests/ -m "not slow"
• ruff check .
• mypy processbehavior (advisory — no new errors)
• Golden masters regenerated deliberately, if the diff touches them

Manual validation:

• short_series_sampling.py reproduces the simulated results reported in Series length (T) as an axis in design-state admissibility — and a saturation question #114.
• short_series_bands.py extends the MR̄/d₂ sampling study to every n from 3 through 30 with independent RNG streams per n, 100,000 replicates, and Monte Carlo standard errors.
• short_series_bands.py checks agreement with the original table within Monte Carlo error.

Notes

The scripts intentionally do not propose an adequacy criterion or band edge. They provide the sampling arithmetic for the maintainers' methodology discussion.

[cnicholas]

Verified locally against the 0.2.1 wheel: short_series_sampling.py reproduces every number in #114 and short_series_bands.py reproduces the n=3–30 table, with its own agreement check against the original within Monte Carlo error. Self-contained, reads d2 from the library, not wired into CI by design.