BREAKING CHANGE: v0.0.5 -> chaintools antirepeat x117 speedup impl

[alejandrogzi]

Full refactor of the antirepeat tool. On a ~200 MB soft-masked chain file with
.2bit reference/query it previously took ~12 h; it now completes in minutes, with
byte-identical output.

Root cause

antirepeat used a lazy .2bit access path: for every chain below --no-check-score
it re-read the sequence from disk and, per fetch, the twobit reader scanned the
chromosome's soft-mask/N block list from the start (a linear skip_while, no binary
search). On repeat-masked genomes — hundreds of thousands of mask blocks per chromosome —
each of the ~millions of per-chain fetches re-scanned a large prefix, and a cost that
grows with chromosome length dominated the run. Parallelism was also off by default.

Performance

• Preload sequences into memory once. .2bit reference/query are now fully decoded
  into memory at startup (the soft-mask/N scan is paid once per chromosome instead of once
  per chain), so every per-chain access is an in-memory lookup. This is the single biggest
  win and turns the ~12 h run into minutes.
• Load only referenced chromosomes. A cheap header-only pre-scan of the chain file
  loads just the sequences it references, bounding peak memory on fragmented assemblies
  (stdin input falls back to loading everything).
• Parallel by default. The parallel feature is now part of the cli build, so chains
  are filtered across all cores out of the box; the previous --threads startup error
  (when built without the feature) is gone.
• Zero per-chain allocation. The filter now borrows chromosome slices directly instead
  of copying each chain's span; minus-strand queries are reverse-complemented on the fly
  during the walk rather than copied and reversed. On large-span, repeat-driven chains this
  is a further ~9.6× faster and ~2× lower peak memory in benchmarking.
• Fused filters. The degeneracy and repeat-mask filters now share a single pass over a
  chain's aligned blocks.
• I/O tuning. Larger input read buffer (1 MB) and larger parallel batches.

Changed

• AntiRepeatEngine::chain_passes no longer takes a SequenceCache; sequence access is now
  through the new SequenceResolver::chromosome() borrowing accessor.
• The score tool also benefits from the in-memory .2bit preload, as it shares the
  sequence resolver.

Notes

• Output is verified byte-identical to v0.0.4 across plus/minus strands, soft-mask and N
  content, gzip input/output, and any thread count, via a randomized differential test and
  full old-vs-new output diffing.