Add DIMTEST (Stout 1987 / Nandakumar & Stout 1993) essential-unidimensionality test

[seonghobae]

Implements DIMTEST (Stout, 1987, as refined by Nandakumar & Stout, 1993) — the confirmatory test of essential unidimensionality — as iteration 27 of the paper-implementation series. Stacked on #257 (ci_classify).

Scope (spec-verify: REDUCED-SCOPE APPROVED)

• Caller-supplied confirmatory AT1/AT2 partitions (equal length >= 4, disjoint); PT = complement.
• Group examinees by raw PT score; discard groups with J_k < 20.
• ML (J_k) variance denominators; refined bias-correction denominator S_k^2 = [(mu4_k - sigma_k^4) + delta4_k + 2*sqrt((mu4_k - sigma_k^4)*delta4_k)]/J_k.
• T_L = K^{-1/2} * sum_k (sigma_k^2 - sigma_U,k^2)/S_k on AT1; T_B analogously on AT2; T = (T_L - T_B)/sqrt(2); one-sided p = 1 - Phi(T).
• Explicitly OUT: ATFIND, DIMTEST 2 bootstrap (Froelich & Habing, 2008), polytomous, missing data.

Citation governance

• READ (formulas transcribed): Nandakumar & Stout 1992 ERIC tech report ED351383 (pre-publication of the 1993 JEBS paper); Kieftenbeld & Nandakumar 2015 (original-vs-bootstrap distinction).
• NOT READ: Stout (1987) original; Stout et al. (2001); Froelich & Habing (2008). Code cites "Stout (1987) as described by Nandakumar & Stout".
• No open-source DIMTEST implementation exists to compare against (sirt = DETECT; ltm::unidimTest = modified parallel analysis); oracle is an independent Python implementation.

Evidence

• Pinned oracle (500x18 2D fixture, rho=0.30, seed 20260726): t_l=8.5848469411043151, t_b=3.6579307315481961, t=3.4838558621150524, p=0.000247122791999742, groups_used=8, n_discarded=17. Rust matches at 1e-12 (p at 5e-7 — crate erfc is the Numerical Recipes approx, |err|<1.2e-7; disclosed in tests).
• Exact-null bits fixture: AT columns = distinct bits of 0..31 give exactly zero pairwise covariance -> t_l = t_b = t = 0 exactly.
• 5/5 mutation kills EXECUTED: (M1) sign swap in sigma^2 - sigma_U^2; (M2) J_k-1 denominators; (M3) dropped bias-correction term; (M4) Jmin 20->2; (M5) two-sided p. Each fails the pinned oracle/null tests; suite restored green (12 passed).
• 500-rep Monte Carlo size/power test (#[ignore]) executed: size < 0.10 at nominal 0.05, power > 0.8 for the 2D alternative.
• Python: TestDimtest 5 passed; validate-before-cast (complex laundering, exact 0/1, index range/overlap) before any dtype conversion.

Adversarial impl-review outcome

• Round 1 (FINDINGS, 2 MAJOR): (1) Python wrapper validated after dtype casts — string responses/indices silently accepted; (2) Rust tests omitted the spec-required per-group oracle intermediates.
• Fixes (93e56e8): dtype-kind checks before any astype coercion (+ string-rejection regression tests); grouping and per-group computation factored into dimtest_pt_groups/dimtest_group_diag, with a new test pinning all 8 groups x 2 subtests x 8 intermediates (J_k, mean, sigma_k^2, sigma_U,k^2, mu4_k, delta4_k, S_k^2, contribution) at 1e-12 against the independent oracle.
• Round 2: CLEAN. Re-verified: refactor preserves numeric operation order; per-group literals match the spec oracle digit-for-digit; bool/int/float still accepted; 13 Rust + 6 Python tests pass.