Add binned quality scores support for modern Illumina platforms

[joshfactorial]

Modern sequencers (NovaSeq 6000, NextSeq 1000/2000) quantize per-base quality into a small set of discrete bins instead of emitting the continuous Q0-Q40 range. This adds a binned_quality_bins option to gen-seq-error-model so users can train a model that snaps observed quality scores to a configurable bin set; gen-reads then emits only those bin values when sampling from the model.

• New optional binned_quality_bins: [..] YAML field on the gen-seq-error-model config. Validated for non-empty, < 94, and excludes 31 (encodes to '@' under Phred+33 and would corrupt FASTQ output). Sorted and deduped at parse time.
• accumulate_qual now snaps each decoded Q-score to the nearest bin during count accumulation (ties round down), so seed, transition, and global counts are all in bin space; error_rate naturally reflects what reads will look like.
• QualityScoreModel::from_counts gains an is_binned parameter and now sets the previously-hardcoded binned_scores field accordingly. Adds a new InvalidConfiguration error variant.
• Public accessor SequencingErrorModel::quality_score_model() for tests and consumers that need model metadata.
• Zero-count bins are kept in quality_score_options and fall back to the existing uniform-row behavior, with a warn! listing them.
• README and template_config updated with the new field, validation rules, and suggested bin sets per platform.

No gen-reads changes needed: generate_quality_scores already restricts output to quality_score_options, which is now constrained to the bin set when the model is binned.