v0.6.1

Patch rather than minor: no public API changed (every helper involved is
crate-private), no new flag or entry point was added, and the on-disk format is
untouched — group sizes 3 and 4 were always valid and always reachable via
--interleaved, so this only changes when they are chosen. Archives already on
disk are unaffected, and both cross-release compatibility directions are
unchanged.

Fixed

• Single-cell interleaving is no longer archived as paired. Layout detection
  only ever compared records pairwise, so a 4-member spot whose members share one
  read name — 10x R1/R2/I1/I2 as fasterq-dump and sracha emit it, where the
  members differ only in length= — looked like a mate pair at every offset and
  was recorded as paired. fqxv info reported twice the real spot count, and
  --split wrote two mate files that each interleaved two different slots at two
  different read lengths, with no warning. Detection now cuts the peeked records
  into runs of consecutive same-spot names and infers the common run length (up to
  4), so 3- and 4-member layouts are detected and restored to one file per slot.
  Archives written before this are unaffected: the stream itself was always
  byte-lossless, only the recorded grouping was wrong.
• Interleaved SRA deflines are detected as paired. @RUN.5 5 length=150 repeats
  the spot name on both mates, and its leading digit was read as a Casava mate
  number — so both mates looked like member 5, and since members must be distinct
  that vetoed the pairing and the stream was archived single-end, leaving --split
  unable to restore the mate files. A leading description digit now counts as a mate
  field only when followed by : (1:N:0:…), which is what distinguishes it from a
  numeric spot name. This is the defline the documented
  sracha get -Z … | fqxv compress - pipeline produces.
• Out-of-step inputs warn instead of silently mis-pairing. With G > 1 member
  assignment is positional, so a stream missing one member mid-file shifts every
  later read into the wrong mate file while the read count stays a clean multiple of
  G — invisible to the spot-multiple check and to every CRC. Compression now
  compares the members' names within each spot and warns with the count and the
  first offending spot. This is the shape fasterq-dump --split-files produces for a
  spot whose READ_LEN slot is 0: it writes no record at all for the empty slot.
• Two FASTQ shapes SRA conversion produces now have container-level tests, having
  been covered only at the codec level: a zero-length read inside a G > 1 spot
  surviving --split in its own slot, and . plus the full IUPAC ambiguity set and
  lowercase bases round-tripping (SRA's 4na→text map is .ACMGRSVTWYHKDBN, so
  converted FASTQ carries more than N).