feat(quant): fill in the mapping-output fields SAM allows to be absent (QUAL, M5/UR, MC/MQ, ZW, @RG, unaligned reads)

[BenjaminDEMAILLE]

Second of three, stacked on #1141. That PR fixes the statements the mapping output makes that are false (a synthesized CIGAR, AS describing a different alignment, MAPQ pinned to 1). This one fills in the fields SAM allows to be absent, which is a weaker claim and deliberately a separate decision: QUAL may be *, and M5, UR, MC, MQ, ZW may simply not be there. The file was still less useful than it had to be, and salmon already had every value in hand.

GitHub cannot point a base branch at a fork, so the diff shown here includes #1141 until that merges. The two commits on top of it are the content of this PR (~1.0k lines).

What gets written

Field	Before	After
QUAL	* / 0xff	The FASTQ's Phred+33, reversed alongside SEQ on the reverse strand. Not behind a flag: whenever the input carries qualities, the record does. A quality string whose length disagrees with the read is dropped rather than written (that mismatch is what makes a record unreadable) and warns once.
@SQ M5	absent	The MD5 of each reference, which is what ties a BAM to the exact bases it was made against and what CRAM needs. Where indexing had to replace a base, the checksum hashes the N that was there (using ambiguous.bin from #1141), so the header names the reference the user supplied. Matches samtools dict.
@SQ UR	absent	The index the bases came from, as a file:// URI, the shape samtools writes.
MC / MQ	absent	The mate's CIGAR and mapping quality, which only became meaningful once the CIGAR was real and MAPQ was not a constant. A mate-aware tool no longer has to sort the file to find them.
ZW	absent	The placement's equivalence-class weight, which salmon already computed and nothing else in the file reports.
@RG + RG:Z	absent	Via --rgLine, in the spelling bwa and STAR users already type (tab-separated KEY:value, escaped \t accepted), parsed before any mapping work so a malformed line fails immediately rather than after the run.
@HD	VN:1.0 SO:unknown	VN:1.6 SO:unsorted GO:query, plus a @CO stating the one ordering guarantee. SO:unknown was not false, but it hid a property the output does have: records for one fragment are contiguous, so a reader can pair mates in one streaming pass instead of sorting first.
unmapped reads	absent	With --sampleUnaligned: FLAG 0x4, no reference, no position, sequence and qualities intact.

--writeQualities becomes accepted-and-default rather than warned-as-unimplemented. --sampleOut stays unimplemented and still warns.

--sampleUnaligned, and the orphan mate

A BAM containing only the reads that mapped is a record of one filtering decision, not of the experiment, and the reads it drops are exactly the ones anybody debugging a low mapping rate wants. --sampleUnaligned no longer requires --sampleOut; on its own it adds those records to --writeMappings/--writeBam, and warns when there is no mapping output for them to go into.

It also covers an orphan's unmapped mate, which is a read that did not align just as much as one whose whole fragment failed. The mapped mate carries MATE_UNMAPPED, so without the mate's record the file described a read it did not contain. The mate is written at the mapped mate's position so a sort keeps the pair together, and the two records point at each other. Without --sampleUnaligned the mapped record keeps RNEXT '*', since naming a mate that was not written is worse than admitting there is none.

Cost

M5 means hashing the whole transcriptome at startup, a few hundred megabytes of MD5. It is one independent hash per reference, so they run in parallel: 0.16s to 0.08s on a 44 MB transcriptome. QUAL and the extra tags are per-record work already dominated by realization; the --writeBam figures in #1141 (1.94s / 3.28s per 1M fragments) are measured with everything in this PR present.

MD:Z is the largest single contributor to file-size growth and is redundant with the reference, so making it optional is the obvious knob if that matters; not done here.

Verification

Against samtools 1.24, on 20k unique transcripts and 5k genes x 4 isoforms (1M fragments each), human chr22 (hg38) + GENCODE v44, and real reads (DRR028171, HiSeq 2500, 457366 pairs of 151 bases):

• samtools fastq round-trips the whole real library back out of the BAM: all 457366 reads returned, not one sequence or quality string altered. This is the strongest check here, since it holds SEQ/QUAL orientation to account on every reverse-complemented record.
• @SQ M5 matches samtools dict, and matches the input FASTA even for the transcript containing an N.
• samtools quickcheck passes; SAM output round-trips through samtools view -b.
• Integration tests cover the header (M5, UR, @RG), the tag set, the read group on every record, the orphan mate, and that unaligned records appear only when asked for.
• Full workspace suite green.

Stack

1. fix(quant): make each --writeMappings/--writeBam record describe the alignment it reports #1141 the false statements (real CIGAR, NM/MD, AS, MAPQ, ambiguous.bin).
2. this PR the permitted omissions (above).
3. feat(quant): --spliced, mapping output in genome coordinates (stacked on #1141, #1142) #1123 genome-coordinate output (--spliced), parked pending the demonstrated-need discussion there.