Add misc/benchmark_truth.py: score both modes against a truth set

[wdecoster]

Refs #23. The harness only — you run it where HG002 lives.

What it does

Runs --mode sensitive and --mode fast over a truth set's loci and scores both against the truth allele lengths, rather than against each other. Every number quoted for --mode fast so far has been agreement with sensitive, which assumes the alignment path is right; this separates "do the modes agree" from "which is closer to the truth".

Design choices worth checking

• The target BED comes from the truth VCF's own coordinates. Both sides then use the same REF interval, so STRdust's RB is directly comparable to len(ALT) - len(REF) from the truth record. Scoring against a separate repeat catalog would reintroduce the boundary mismatch that makes such comparisons ambiguous — the same mismatch behind the neighbour-locus disagreements in Add --mode fast: genotype straight from the existing alignment #25.
• Loci outside the benchmark's confident regions are skipped, not silently scored, via --confident-bed.
• QUICKREF loci count as a reference-length call, not a missing one — otherwise the fast path's cheapest calls would vanish from the denominator.
• Results are stratified by how far the truth allele departs from the reference (reference / 1-10 / 11-50 / 51-200 / >200 bp), which is where the two modes are expected to part company.
• No truth-set URLs are hardcoded: the GIAB FTP layout changes between releases, so the docstring points at the site and you pass the paths.

Verification

Feeding a STRdust VCF back in as the truth set is an identity check — --mode sensitive then scores 100% exact across every stratum, which is what confirms the coordinate handling and allele pairing are right. On a 30x ONT sample over 300 loci:

mode       stratum    loci_scored  exact_pct  within_5bp_pct  median_abs_diff  cpu_seconds
sensitive  all        300          100.0      100.0           0.0              71.2
fast       all        300          26.0       97.0            1.0              0.4
fast       reference  48           95.83      100.0           0.0              0.4
fast       1-10bp     247          12.96      96.76           2                0.4

(That is the fast-vs-sensitive comparison again, since the pseudo-truth is sensitive's output — it is here to show the harness works, not as a result.)

Also exercised: --from-vcfs, --plot, --regions, --max-loci, and the argument validation.

Running it

python misc/benchmark_truth.py \
    --binary ./target/release/STRdust \
    --fasta GRCh38.fa --bam HG002.cram \
    --truth-vcf HG002_TR_benchmark.vcf.gz \
    --confident-bed HG002_TR_benchmark_regions.bed \
    --threads 8 --out-prefix hg002_tr --plot

Standard library only for running and parsing, so it needs no environment beyond Python; matplotlib is imported only for --plot. Start with --regions chr1 --max-loci 2000 to check the truth set parses the way you expect before committing to a whole-genome run.