_write_from_svar (SVAR1) ignores max_mem

[d-laub]

Follow-up from #333.

#333 made _write_from_svar2 honour gvl.write's max_mem by consuming genoray's chunked _find_ranges_chunked stream. The SVAR1 path, _write_from_svar, still ignores max_mem entirely.

Lower priority than the SVAR2 case was, for two reasons:

• different mechanism — SVAR1 uses _find_starts_ends_with_length(..., out=), writing into a caller-provided array,
• no transient amplification — there is no equivalent of the SVAR2 path's three-copies-of-the-payload peak that made SVAR2: gvl.write ignores max_mem and OOMs on whole-contig range materialization at 414k samples #333 fatal.

But the gvl.write docstring now singles out .svar2 as the source that honours max_mem, which documents the SVAR1 gap without tracking it. This issue tracks it.

Related: the SVAR1 max_ends tie-break at same-POS SNP+DEL differs from SVAR2's (SVAR2 is the correct one).