Streaming GenotypeMatrix honors accessible_bed for grm/ibs (#152)

[andrewkern]

Closes #152.

Problem

GenotypeMatrix.from_zarr(..., streaming=..., accessible_bed=bed) accepted accessible_bed but both streaming branches dropped it, so a StreamingGenotypeMatrix carried no accessible mask. grm/ibs then streamed over all variants while the eager path filtered to accessible sites via the mask-filtered .genotypes view -- a silent eager-vs-streaming divergence under an accessible_bed, with no error raised. This is the sibling of #151, which fixed the same drop for HaplotypeMatrix.

Fix

• Forward accessible_bed from both GenotypeMatrix.from_zarr streaming branches into _build_streaming, resolve the mask once over the source's variant bounds, and hand it to StreamingGenotypeMatrix. iter_gpu_chunks already attaches the mask to every chunk, so per-chunk variant filtering happens for free and streaming grm/ibs match eager.
• Factor the shared "resolve mask over source bounds" block into resolve_streaming_accessible_mask, now used by both the haplotype and genotype streaming loaders.
• Correct the stale StreamingGenotypeMatrix docs/error text, which still claimed grm needs .materialize() -- it streams directly.

Audit

Swept the whole streaming surface for this bug class (mask honored on eager, dropped on streaming):

• Loaders: only from_zarr streams; both HaplotypeMatrix (Streaming span-normalization ignores accessible-span metadata (mask / n_total_sites), diverging from eager #151) and GenotypeMatrix (this PR) now forward the mask. from_vcf/from_ts are eager-only.
• Consumers: no stats module reads raw arrays -- all go through the accessible-filtered .haplotypes/.genotypes/.positions properties. Verified per streaming consumer: sfs, genetic_relatedness (Streaming genetic_relatedness: honor the accessible mask (span + variant filtering) #153), grm/ibs, windowed_analysis, and ld_statistics all honor the mask. selection, admixture, diversity, divergence, distance_stats, ld_statistics_genotype, and local_pca have no streaming-matrix input path (eager-only or raise).

Conclusion: no other statistic carried a live instance of this bug. The one residual was a test-coverage gap for the load-time-mask -> streaming windowed_analysis path, closed here with a regression test.

Tests

• Eager-vs-streaming grm/ibs parity with and without an accessible_bed, plus a guard that the mask actually changes the streaming result and a chunk-level check that each streaming chunk is restricted to accessible positions.
• Pin the load-time mask through streaming windowed_analysis.
• Full suite green (test_streaming_kernels, test_accessible_mask, test_relatedness, matrix/streaming loaders); ruff clean.

[nspope]

Looks good---one comment about missing test coverage.

[nspope]

This is good, but doesn't have a case where there's missing data. It should be easy enough to inject missing data (-1) into the vcz --- I think that's important to cover for the parity check between streaming and eager (cover interaction between missing data (both modes) and the mask).