perf: reuse cohort allele counts in pairwise_average_fst

[kshirajahere]

Summary

This improves pairwise_average_fst() by reusing per-cohort allele counts instead of recomputing them for every cohort pair.

Before this change, pairwise_average_fst() delegated each pair to average_fst(), which meant the same cohort query could trigger repeated snp_allele_counts() work across many pairs. With N cohorts that grows to N(N-1) allele-count computations.

After this change, pairwise_average_fst() computes allele counts once per cohort for the given parameter set, then reuses those arrays for the pairwise Hudson Fst calculations. average_fst() behaviour is unchanged.

Fixes #692.

Why this matters

pairwise_average_fst() is a core analysis path for exploratory population comparisons. Improving its runtime helps interactive use in notebooks/Colab and makes higher-level workflows more responsive.

Changes

• factor the shared blockwise Hudson Fst logic into _average_fst_from_allele_counts()
• keep average_fst() as the public two-cohort entry point, now calling the shared helper
• precompute cohort allele counts once inside pairwise_average_fst() and reuse them across all cohort pairs
• add a regression test proving pairwise_average_fst() still matches average_fst() for each pair
• add a regression test proving each cohort query only triggers one snp_allele_counts() call within pairwise_average_fst()

Tests run

• ruff check malariagen_data/anoph/fst.py tests/anoph/test_fst.py
• ruff format --check malariagen_data/anoph/fst.py tests/anoph/test_fst.py
• pytest tests/anoph/test_fst.py -q

Notes

• I did not run the full repository test suite.
• pytest tests/anoph/test_fst.py -q passed, with pre-existing warnings from the simulated adir1 fixtures about missing surveillance flags metadata.