Pin statistics across implementations in a parity test suite

[andrewkern]

Summary

Addresses #135. Adds tests/test_implementation_parity.py, a suite that asserts the same population-genetic statistic agrees across the code paths that compute it: scalar functions, FrequencySpectrum, and the windowed scatter, fused CUDA, and python-loop engines.

Targets nsp-multiallelic-trunk (this suite validates the multiallelic work), not main.

Each statistic is described once in a declarative _STATS registry; the parametrized test drives every supported path across four data conditions — clean, missing/include, missing/exclude, and multiallelic — and asserts it matches the scalar reference. 16 statistics, ~180 cases: 149 passed, 12 skipped, 19 xfailed. GPU-only (skipped without CUDA).

This is pure test infrastructure — no library code changes. Divergences between paths are recorded as strict=True xfails in _XFAILS with a mechanism-level reason; a (stat, path) may carry several condition-scoped rules. Fixing any divergence removes its rule and flips the cell to a pass.

Divergences this suite currently pins

1. fst_wc (fused; all conditions) — the fused kernel was not converted to the per-allele one-vs-rest Weir-Cockerham ANOVA the scalar/python-loop path now uses.
2. theta_w, tajimas_d (fused; missing/include) — the fused kernel's per-site sample-size handling under missing data still diverges from the scalar.
3. tajimas_d, normalized_fay_wu_h, zeng_e (fs and scatter; missing/include) — the Achaz neutrality-test variance uses a modal / full-sample size rather than per-site valid counts.
4. pi, theta_w, tajimas_d, segregating_sites, fay_wu_h, normalized_fay_wu_h, zeng_e (fs; multiallelic) — FrequencySpectrum builds a per-allele SFS whose segregating-class count and pi differ from the scalar's per-allele counts.
5. da (scatter; missing/exclude) — the within-population pi terms use a different site set than the scalar (dxy agrees).

Notes

• The fused engine only runs under missing_data='include', so it is skipped (not failed) under the exclude condition.
• Two-population fixtures use an even haplotype count per population, which fst_weir_cockerham requires for its diploid pairing.

[nspope]

did you base off of #132 [that is, merge target is nsp-multiallelic-trunk]? I think that fixed the singleton and da and fst issues, at least

[andrewkern]

shit, I did not. not me pull in those changes to this branch

[nspope]

i'd just rebase on nsp-multiallelic-trunk, that's where i'm doing all the multiallelic stuff so I don't break main in the meantime

[andrewkern]

okay this should be ready to go now @nspope