deps: noodles 0.113 -> 0.115 and noodles-bgzf 0.49 -> 0.51 (local thread pools, with_worker_count)

[BenjaminDEMAILLE]

noodles moved fast in the last two weeks and the pins here are two minors behind.

Crate	Pinned	Latest	Released
noodles	0.113	0.115.0	2026-08-03
noodles-bgzf	0.49	0.51.0	2026-08-03

Pre-1.0, so minors are breaking by convention: this needs reading, not a blind bump.

Why bgzf 0.50 / 0.51 specifically matter here

• 0.50 — multithreaded readers and writers moved off rayon's global pool onto exclusive local
  pools, and the deprecated with_worker_count methods came back as the supported way to size them.
• 0.51 — worker count is now synced to the number of managed buffers instead of rayon's global
  pool, so merely constructing one no longer initialises that pool.

This codebase leans on rayon for alignment (src/lib.rs and friends). A BGZF component that
quietly borrows the global rayon pool competes with the align workers for the same threads; a
component with its own sized pool does not. That is the interesting part of the upgrade.

The adjacent opportunity

src/io/bam.rs:48 builds a plain single-threaded bgzf::io::Writer for every BAM sink
(:60, :374). BAM output is therefore serialised behind one compressor while alignment runs on N
threads. With 0.50+ the multithreaded writer is usable with an explicit, non-global worker count,
which is exactly the shape needed here.

Caveat that must be checked, not assumed: BGZF block boundaries and therefore output bytes can
depend on the writer's buffering, so a multithreaded writer must be shown to produce a
byte-identical file (or a documented, deliberate change) at every worker count.

Checklist

• Read the noodles changelogs 0.113 → 0.115 for fastq/sam/bam, note breaking changes
• Bump noodles and noodles-bgzf in separate commits from any behaviour change
• Full test suite + the faithfulness benchmark, before/after
• Separately evaluate the multithreaded BGZF writer with an explicit with_worker_count
• Prove BAM bytes identical across worker counts, or document the change
• Also on the table while here: bstr 1.13, rustc-hash 2.1.3, memmap2 0.9.11 (all
  compatible with existing ranges); dashmap 7 is a release candidate, stay on 6