Unify the biallelic definition across GenotypeMatrix loaders and make biallelic restriction evident with warning

[nspope]

Background

Two related problems. (1) Several entry points restrict to biallelic sites without telling the user,
so a biallelic-only statistic run on multiallelic data silently returns a partial result. (2) The
loaders disagreed on what "biallelic" even means, and on how to build a dosage -- so the same data
produced different genotype matrices depending on which loader you used, and none of that was
caught by the test suite because nothing compared the loaders.

This PR adds a shared BiallelicOnlyWarning, unifies the biallelic definition and the dosage rule
across all three GenotypeMatrix loaders (from_vcf, from_zarr eager, from_zarr streaming) and
from_haplotype_matrix, and adds a cross-loader parity harness.

The unified definition

• Biallelic = at most two distinct present alleles (label / reference independent). A {0,2} site,
  or a reference-absent {1,2} site (possible after subsetting), is biallelic. This replaces the
  narrower "codes in {0,1}" notion that wrongly dropped genuinely-biallelic non-{0,1} sites.
• Dosage = count of the chosen alt allele: allele 0 is reference, alt = highest present allele > 0,
  with a sentinel (out of range) when no alt is present so a monomorphic site counts to 0. This
  replaces summing allele indices (wrong for {0,2}) and gt > 0 (folds distinct alts; breaks on
  reference-absent {1,2}).
• {0,1} input stays byte-identical, including a monomorphic-derived {1}-only site (dosage 2). This
  is what forced "allele 0 = reference" rather than "highest present allele" in the alt choice.
• Both loaders retain monomorphic and all-missing sites (they load faithfully); dropping
  uninformative sites is the job of apply_biallelic_filter, not the loaders. The diploid stats
  tolerate such columns (grm excludes them via its polymorphic mask; pca/ibs guard every division).

The source of truth for biallelic sites is _biallelic_and_alt(ac) in genotype_matrix.py.

What changed

• _warnings.py: BiallelicOnlyWarning(UserWarning) plus _warn_biallelic_only(n_dropped, *, context,
  stacklevel): coerces the count to int, no-ops at 0, emits a message leading with the context and
  dropped-site count.
• GenotypeMatrix.from_haplotype_matrix: onto the shared definition. Counts the alt allele instead of
  summing indices; kepps {0,2}/{1,2}/monomorphic (previously {0,2}/{1,2} were dropped and index-sum
  gave impossible dosages like 2+2 -> 4); drops >= 3-allele sites + warns. {0,1} bit-identical.
• GenotypeMatrix.from_vcf: moved off scikit-allel's is_biallelic_01() onto the shared definition.
  keeps {0,2}/{1,2}/monomorphic (is_biallelic_01 dropped all three); counts the alt allele; builds
  the allele-count matrix with numpy so scikit-allel's library footprint is now purely VCF parsing
  (read_vcf). Drops >= 3-allele sites + warns.
• GenotypeMatrix.from_zarr (eager AND streaming, via build_genotype_matrix): onto the shared
  definition. Both zarr paths used a raw ploidy-sum with no classification, so a {0,2} site or any
  multiallelic site produced impossible dosages, silently, and disagreed with from_vcf. Now counts
  the alt allele. A >= 3-allele site has no valid 0/1/2 dosage, so its whole row is recoded to -1
  (present but fully missing) -- this preserves the row/chunk alignment the streaming path relies on
  (from_vcf drops the site instead). The warning is surfaced by the callers: the eager path with the
  exact count, the streaming path once per load on the first affected chunk.
• admixture.patterson_d: warn with the count of >2-allele sites it already drops (num=den=0), once
  per top-level D computation (patterson_d / moving_patterson_d / average_patterson_d). D unchanged.
• GenotypeMatrix.apply_biallelic_filter: docstring corrected. It does no allele classification (the
  matrix is biallelic by construction) -- it drops monomorphic (non-segregating) sites. Name kept
  for backwards compatibility; a rename is deferred as a follow-up.

Behaviour

• {0,1} input is byte-identical across every loader, silent.
• {0,2} / reference-absent {1,2} / monomorphic sites are now KEPT (and correctly dosed) by all
  loaders; previously from_vcf dropped them and the other paths mis-handled them.
• A >= 3-allele site: from_vcf and from_haplotype_matrix drop it (fewer variants); from_zarr recodes
  its whole row to -1 (present but fully missing). All paths emit one BiallelicOnlyWarning.
• All three GenotypeMatrix loaders now produce byte-identical dosage matrices from equivalent
  biallelic data, and stay consistent on multiallelic data.
• The warning is silenceable by category (from pg_gpu import BiallelicOnlyWarning;
  warnings.filterwarnings("ignore", category=BiallelicOnlyWarning)).

Not in scope / deferred

• LD / moments-LD paths keep the old "codes in {0,1}" notion by design (they are on their way to a
  separate multiallelic-correct LD rewrite): haplotype_matrix.apply_biallelic_filter (a moments
  is_biallelic_01 reimplementation), the moments-LD biallelic filter + alt construction, and the
  Rogers-Huff matrix dosage. These three call sites are documented in the LD subproject plan so the
  reconciliation is not lost. rogers_huff itself restricts to biallelic and will use this warning
  there.
• apply_biallelic_filter is misnamed (it drops monomorphic sites, not a biallelic classification).
  Docstring fixed here; a rename or a clearer alias with soft-deprecation is a public-API change
  deferred as a follow-up issue.

Verification

• tests/test_biallelic_only_warning.py: the class and helper (subclass, count/context message,
  numpy-scalar count, no-op at 0, silenceable); the shared _biallelic_and_alt definition and alt
  choice; from_vcf keeps {0,2}/{1,2}/monomorphic and drops the triallelic site with correct
  dosages; from_haplotype_matrix bit-identity on {0,1}, alt-code independence, reference-absent
  {1,2}, and the one-missing-haplotype -> fully-missing (-1) case; degenerate-column downstream
  safety (grm and pca_dosage stay finite on explicit monomorphic and all-missing columns);
  patterson_d warns once.
• tests/test_loader_parity.py (new): drives from_vcf, from_zarr (eager), and from_zarr (streaming)
  from a single call_genotype block and asserts they agree byte-for-byte on biallelic data, stay
  consistent on multiallelic data (zarr recodes to a -1 row where from_vcf drops), and that the
  streaming warning fires exactly once across multiple chunks. This is the cross-loader coverage
  that was missing when the three loaders silently diverged.