SegPE v0.2.0

Rebuilt around byte arenas and a precomputed adapter variant index.

Performance, 2.5M read pairs (1.67 GB), 384 anchored 16 bp adapters,
against v0.1.16 on the same data and machine:

  Windows x86-64, 32 threads   2.31 s -> 0.68 s
  Linux x86-64, 32 threads     7.85 s -> 0.41 s
  Linux aarch64 (GB10), 20     6.60 s -> 0.34 s
  macOS aarch64 (M5), 10 cores           0.33 s

Output is byte-identical across all four platforms. Compressed input:
BGZF 0.48 s, ordinary gzip 1.45 s. Against cutadapt 5.2 on the same task,
7.43M vs 1.95M reads/s with zero conflicting classifications.

Behaviour changes that affect existing pipelines:

  - --error-tolerance now means edit distance <= e. It was a score cutoff
    that admitted fewer than e/2 mismatches, so the default e=1 accepted
    only exact matches. Runs will classify more reads than before.
  - Multi-member gzip is read in full. v0.1.x stopped at the first member,
    so bgzip input was silently truncated to roughly its first block.
    Re-run anything processed from bgzip files with v0.1.x.
  - Write failures now reach the exit code instead of being logged while
    the run reported success.
  - Ambiguous demultiplexing is counted and warned about; --ambiguous
    controls whether those reads are kept, separated or rejected.

New: BGZF block-parallel decompression, bzip2 and zstd input, --max-mem,
optional wgpu GPU backend (off by default, and slower end to end).

Dependencies cut from 15 to 8. 25 unit tests, run on x86-64 and aarch64.

Support LowQualityControl and improve code

support single-ends, and speed up processing

Speed up 8x

first public version