feat: add LD pruning and admixture data preparation

[31puneet]

Overview

Fixes #1049

• Adds ld_prune — a standalone LD pruning function using iterative Rogers-Huff r² via allel.locate_unlinked(),
• Adds prepare_admixture — chains LD pruning → optional SNP downsampling max_snps → PLINK binary export (.bed/.bim/.fam)
• Supports pre-filtering via both individual downsampling (cohort_size) and SNP downsampling max_snps
• 10 new tests across both modules, all passing

Impact

• Users can now perform LD pruning independently for PCA, GWAS, or
  other analyses that assume linkage equilibrium
• Users can prepare ADMIXTURE-ready PLINK files directly from the API
  without manual preprocessing

[31puneet]

Hi @jonbrenas, this PR is ready for review. I’d appreciate your feedback. Thanks!

[31puneet]

Hi @jonbrenas updated the PR
ld_prune returns an xr.Dataset with caching, and prepare_admixture wraps it with random downsampling + PLINK export following to_plink.py conventions. Tested on real Ag3 data (3L, AG1000G-BF-A) — ~10.9M variants → 658K after pruning → 50K with max_snps, valid PLINK output verified with bed_reader.