compair 0.3.0

tl;dr moves to seqair-types 0.4.0, so a Read always has a known strand. The batch and pairs kernels renormalise with SIMD, with scores unchanged.

Breaking

• seqair-types 0.4.0: the re-exported Strand has no Unknown, so a Read always carries a known strand and Error::UnknownStrand is gone.

Performance

• The batch and pairs kernels renormalise in vectors at each strip crossing instead of lane by
  lane, and compare row numbers against a lane of read lengths instead of rebuilding two masks per
  row. Scores are unchanged. Little difference on M4 or AVX-512; with fearless_simd 1.2 (pick it
  up with cargo update -p fearless_simd) the strip kernel paths are about 6% faster than 0.2.1
  in target-cpu=native builds on AVX-512. Thanks to Dylan Costo
  for the AVX-512 measurements (#84).