add AVX2/AVX-512 to ksw_extd2 with runtime dispatch

[nh13]

ksw_extd2 (dual-gap banded extension) is minibwa's main base-alignment cost for long reads (~50–66% of HiFi/ONT CPU) and was 128-bit only. This adds AVX2 and AVX-512BW implementations of the gap-fill path, selected at runtime via __builtin_cpu_supports so a single binary auto-uses the widest available ISA and falls back to SSE4.2/NEON otherwise. Only the banded approx+cigar path is widened; every other flag combination calls the original kernel.

Output is bit-identical to the SSE4.2 build on all four HG002 zenodo sets vs GRCh38, across sse/avx2/avx512 (~4.4M alignment records, 0 diffs).

End-to-end on a c7i (Sapphire Rapids), user CPU vs SSE4.2:

set        readlen mean/median   avx2    avx512
WGS-1M     151 / 151             1.00x   1.00x
HiC-1M     151 / 151             1.00x   1.00x
SBX-HG002  241 / 227             1.00x   1.00x
HiFi-10k   18881 / 18047         1.15x   1.19x
ONT-10k    19810 / 18180         1.11x   1.14x

The win needs wide-band DP — long, divergent reads where the gap-fill bands are wide; it is absent on short/mid reads and on low-divergence data of any length (confirmed by simulation: flat with small indels at every length, ~1.09x only at 16 kb with large indels). MINIBWA_EXTD2_ISA=sse|avx2|avx512 forces a tier for testing.

[lh3]

I like the way AVX2/AVX512 is implemented but this will considerably increase the burden of maintenance in the long term. Given that the short-read performance remains the same and minimap2 still dominates long-read alignment, I will not merge now. I will keep the PR open until there is a real need.

[nh13]

I agree there's a tension between maintenance burden and trying to make minibwa as fast as one can, supporting both long and short reads, as well as generally trying to favor simplicity. Notheless, the small speed improvements compound, especially if we do them earlier in development, as more executions of minibwa will benefit. I am hopefull that coding agents can help with the maintenance, is that something you'd be open to?

[nh13]

Independently reproduced on a fresh c7i.4xlarge (Xeon 8488C), this time as wall-clock of the mapping phase rather than user CPU, five reps per tier interleaved with order reversal, -t 16, SAM to /dev/null:

             avx2    avx512   rep spread
  HiFi-10k   1.123x  1.164x   0.28-1.43%
  ONT-10k    1.109x  1.134x
  WGS-1M     1.000x  1.000x   0.14-0.25%

Same shape as the user-CPU numbers above, slightly lower as expected for a wall-clock metric that includes the unwidened parts of the run. SAM output byte-identical across all three tiers on HiFi and WGS, re-confirming the bit-identity check. MINIBWA_EXTD2_ISA made this easy — one binary, three tiers, so the comparison carries no build or link differences.

Two notes for whoever picks this up:

The short-read rows were measured with -x sr, which #63 shows is currently the same preset as -x adap. The numbers are correct as measured; the label is what is wrong. If #63 lands and sr becomes genuinely non-adaptive, those rows want re-taking.

#64 speeds up the 128-bit ksw_extd2/ksw_extz2 by 5-11% depending on read length. This PR's dispatcher calls the reference kernel for every flag combination except banded approx+cigar, so the two compose, but the ratios above are against an unimproved baseline and would shrink if #64 lands first.