Align in-memory splice-junction records

[justinblethrow-cloud]

Summary

• Align every typed field in the in-memory Junction record.
• Pad each record from 23 to 24 bytes so records remain aligned in arrays.
• Back the byte-addressed record buffer with vector<uint64> to guarantee suitable base alignment.
• Add compile-time alignment regression checks.

Problem

The current 23-byte record places 32-bit counters at offsets 11 and 15 and 16-bit overhangs at offsets 19 and 21. Junction then exposes typed pointers to those fields and dereferences them. UBSan reports these stores during normal spliced-read mapping, beginning in ReadAlign_outputTranscriptSJ.cpp.

This is an internal runtime buffer change only; no genome-index or output-file format changes.

Validation

• extras/tests/scripts/testJunctionAlignment.sh passes; its static assertions fail against unmodified 2.7.11b on the storage type, field offsets, and record stride.
• A full ASan/UBSan build succeeds.
• Spliced-read mapping no longer reports junction-record alignment errors.
• SAM records, SJ.out.tab, and ReadsPerGene.out.tab are byte-identical to unmodified 2.7.11b.