v9.1.0 — nanopore basecaller: 70% → 74% on real R9.4

A measured accuracy improvement to the Level 7 nanopore basecaller (pure NumPy, no AI),
on real captured R9.4 signal (E. coli, n=36, vs the production Guppy basecall), plus a
robustness bug sweep.

What changed

• Iterative per-read rescaling: a first Viterbi pass tells us which k-mer each
  event is; the per-read scale is then refit (robust least squares) to the actual
  levels of those k-mers, correcting the read's composition bias.
• Tuned transitions (p_stay 0.42 → 0.50): the old value over-penalised the STAY
  states that over-segmentation produces.
• Measured, not assumed: per-k-mer emission noise, drift correction and EM model
  re-estimation were all tried and dropped (they didn't help) — the fixed classical
  model is at its ceiling. Clean ablation.
• Robustness hardening: bug sweep of the new module — fixed two edge-case bugs
  (empty-signal path handling; inf/nan inputs no longer crash the rescaler) and added
  adversarial regression tests.

Result

Mean ~74.5%, median ~74%, range 68–82% (was 70.3% in v9.0). Reproduce it:
python tools/bench_basecaller.py. Still the classical R9-era route, still far below
neural Dorado's ~99% — the honest ceiling of a fixed model; R10 remains out of scope.

Tests

525 passing.