test: gvl.concat has no svar2 coverage — add an .svar2 concat fixture

[d-laub]

Split out from the gvl.concat work (#334, PR #339). Deferred deliberately: an .svar2 fixture needs test infrastructure the concat plan never scoped.

What is missing

gvl.concat implements the .svar2 backend (_concat_svar2_ranges in python/genvarloader/_dataset/_concat.py), but no test exercises it. The concat test suite covers PGEN/VCF on both axes, .svar on both axes, and tracks/annot tracks — but the entire svar2 branch ships untested.

The blocker is fixture infrastructure: tests/_builders/case.py's Case dataclass has pgen_path and svar_path but no svar2_path, so there is no session-scoped .svar2 store to shard and merge.

Why it was not treated as blocking

A review verified the svar2 merge code against the real on-disk format rather than against tests, and found it consistent:

• filenames (vk_snp_range.npy, vk_indel_range.npy, dense_snp_range.npy, dense_indel_range.npy, sample_cols.npy, svar2_meta.json)
• the raw-memmap vs np.save split — the four range arrays are written headerless via gather_fixed/link_or_copy_buffered; sample_cols goes through np.save, matching the reader's np.load
• the 16-byte record stride and (R, S, P, 2) C-order slot space
• the four ["shape"] updates in svar2_meta.json, with "ploidy" preserved
• the reader derives R from dense_snp_range.shape[0] and S from vk_snp_range.shape[1], both of which the merge sets correctly

Worth noting the region-axis dense_* gather also fixes a latent block-concat bug that was present in the original plan's np.concatenate approach.

So: consistent-by-inspection, but unverified by execution.

What to do

1. Add an .svar2 store to the shared synthetic case (svar2_path on Case, built alongside the existing .svar).
2. Add concat fixtures sharding it by region and by sample.
3. Assert against a single-shot gvl.write oracle, mirroring the existing svar tests.
4. Include an interleaved sample split (e.g. [s0, s2] + [s1]). This matters: the plain fixtures split samples alphabetically, so after re-sorting they collapse to a block layout — the exact case a wrong slot ordering would also pass. Only an interleaved split proves the order= handling is right for svar2.

Deferred minors from the same review

Worth folding in if someone is already in this code:

• svar_link/svar2_link are copied verbatim into merged metadata, but relative_path resolves against the dataset dir, so it is stale for a merged dataset written elsewhere. Resolution silently falls back to absolute_path, so tests pass — a later tree move breaks it. Recompute the link against the new destination.
• _gather_svar_offsets has unused n_src_slots/n locals, re-reads each input's offsets file once per plane (2x IO), and np.stack(planes).tobytes() adds two full-size copies on top of the permitted materialization. Writing each plane's bytes separately would halve peak memory inside that carve-out.