dev-deps: add a benchmark harness (divan or criterion) — there is no benches/ today

[BenjaminDEMAILLE]

There is no benches/ directory and no benchmark harness in Cargo.toml. Several open dependency
questions (SIMD crates, packed-seq, external sort, sufr vs caps-sa, #162) all start with
"measure first", and each currently has to invent its own measurement. One harness, once, unblocks
all of them.

Candidates

Crate	Version	License	Note
divan	0.1.21	MIT OR Apache-2.0	Simple API, generic benchmarks, allocation counting
criterion	0.8.2	Apache-2.0 OR MIT	Incumbent, statistics and regression reports, heavier

Both are dev-dependencies only, so nothing ships in the binary and the supply-chain argument in
CONTRIBUTING.md is much weaker than for a runtime dep.

What to benchmark

The paths the open questions actually care about:

1. genomeGenerate suffix-array construction (src/index/sa_build.rs), in-mem and ext-mem arms
2. Seed extension / find_stop (src/align/simd_scan.rs)
3. FASTQ decode, plain and gzipped (src/io/fastq.rs)
4. BAM write and coordinate sort (src/io/bam.rs)
5. End-to-end align on a fixed small fixture, as a coarse regression signal

Constraints

• Benchmarks must not gate CI correctness; a slow machine must never fail the build.
• Numbers are machine-dependent. Useful as A/B within one machine and one session, not as absolute
  thresholds checked into tests.
• Keep the fixture small enough to run locally, large enough not to be dominated by process startup.

Checklist

• Pick one harness, not both
• Add the five benches above
• Document in CONTRIBUTING.md how to run them and how to report an A/B in a PR
• Link this issue from the dependency issues that are gated on measurement