PyData prototype genetics method implementations

[eric-czech]

This issue tracks progress on method implementations with a focus on those mentioned in #16 (or things needed by these methods).

Progress:

• axis_intervals
  • This identifies variants within BP windows
  • Very similar to hail.locus_windows
• maximal_independent_set
  • Somewhat similar to hail.maximal_independent_set but with chromosome partitioned sequential algorithm for compatibility with PLINK/skallel results
    • My rationale for doing this instead is that users would be less skeptical if results were identical to other tools, rather than rolling a more scalable but less credible heuristic for this from the start
• ld_matrix
  • Very similar to hail.ld_matrix
  • There are CPU and GPU implementations for this now
• ld_prune (PyData prototype LD prune implementation #26)
• GRM/RRM
  • GRM
    • Center variants (in rows) by subtracting nanmean
    • Divide by binomial variance for variant under HWE (Patterson 2006)
    • Compute XX^t (for X as n_samples by n_variants)
  • RRM
    • Same as GRM except that empirical variance is used rather than binomial variance under HWE
    • This IS pearson correlation up to a constant factor
  • This gist has GRM calc, which uses same scaling as default preprocessor to PCA in scikit-allel
• HWE (https://github.com/pystatgen/sgkit/pull/76)
  • For axis reductions in variant/sample QC as well as PCA normalization
• PC-Relate (PyData PC-Relate Integration #35)
• PCA (https://github.com/pystatgen/sgkit/pull/262)
• LMM and LFM