Drop GWAS variants on uninterpretable-id LD-panel entries

[Yining97]

Some LD panels (ADSP R5) name ~3% of variants with a tag where an allele belongs — chr21:13988152:INS:T whose real alleles are A/AT — so the id no longer encodes its alleles even though the A1/A2 columns are correct. Harmonisation keys on the allele columns and keeps the variant; the id-string LD lookup (.ldFromSketchMatch) then can't find it and aborts (.fmLdFromSketch / kriging prefilter: N variant id(s) not present in the LD sketch panel). R4 has none of these, so only R5 fails.

The panel's alleles are trustworthy but the summary statistics are not: at such a position the GWAS carries no way to say which orientation was measured, so aligning its z-score would be a guess that can sign-flip a real signal. Dropping is the only choice that can't be wrong — the same reasoning #576 already applies to flip-pairs.

• New .ldSketchIdUsable predicate: an id is usable unless it parses to a non-DNA allele (tests for DNA, not the literal INS/DEL; rsIDs and legitimate multi-base indel alleles stay usable).
• .matchAgainstSketch filters .refVariantsFromSketch by it, so a GWAS variant on such a position is dropped at harmonisation (removeUnmatched) — filtered at the call site, not in .refVariantsFromSketch, since the RAISS path indexes into that frame.
• .raissUnsafeToImpute gains it as a third clause. The id-string readers then never request the variant.

Validated on the real ADSP R5 chr21 panel: baseline reproduces the abort; the fix drops exactly the tag-position variants at QC and the run completes, through both direct calls and the real gwas_sumstats_construct.R + fine_mapping.R scripts. Suites pass with 0 unrelated changes; rsID panels unaffected.