Use pairwise-complete allele frequencies for naive-r2 under missing data

[nspope]

Closes #220

_pairwise_ld_core computed D = p_AB - p_A*p_B where p_AB (the joint 1,1-frequency) was summed over the pairwise-complete sample set (gametes valid at both sites) but p_A/p_B were each site's own marginal frequency (summed over all of that site's valid gametes). Under structured missing data those are different sample sets, so D and r2 came out biased -- in one case, a mathematically impossible r2 > 1. missing_data='exclude' never exercised this, since dropping every site with any missingness makes the marginal and joint sets trivially coincide.

p_A/p_B are now conditioned on the pairwise-joint-valid set for each pair, matching the standard 2x2-table formula _tile_counts/_tile_sigma_d2 already use elsewhere in the LD module. pairwise_r2, pairwise_LD_v, and locate_unlinked (the three callers of the shared core) all pick up the fix; the tiled naive-r2 path used by zns/omega (_tile_r2_naive) is rebuilt on the same _tile_counts helper the projection estimator already uses, dropping its now-unnecessary marginal-frequency parameters.

sigma_d2 is untouched: it's a different, deliberately unbiased estimator -- it was just never affected by it, being built entirely from pairwise joint counts with no marginal-frequency vector anywhere in its own construction.

[andrewkern]

this looks right for the haplotype path, but two things to fix here before we merge:

1. we need the same fix on the genotype path
2. mem usage looks like it goes way up. At 8,000 sites the memory pool goes from 2.66 GB to 4.2 GB, from the new sum_i, p_i and p_j m x m arrays. Since joint_n is symmetric, p_j is exactly p_i.T, so it can be a view, not a new array. And p_i and p_AB can be computed in place into sum_i and joint_11.

also looks like the changelog is now in conflict

[nspope]

this is a bit of a cludge, as it relies on cupy replacing 0/0 with NaN and being silent about it. The alternative is to explicitly index on ~zero, but this creates intermediate copies which inflates memory usage (and the whole scheme here is designed to limit memory usage).

[nspope]

take another look @andrewkern ? Memory usage is now not much higher for your example than the original implementation (but see caveat above), and the genotype path is implemented.