fix: derive the MAGSAC reference threshold from sigma_max (unit-independent)

Problem

The pybind find*_ functions in src/pymagsac/src/magsac_python.cpp leave the
MAGSAC reference threshold (interrupting_threshold, magsac.h:55) at its
constructor default of 1.0. This threshold counts "confident" inliers
(residual < interrupting_threshold), which drives the adaptive iteration count:

last_iteration_number = log(1−conf) / log(1 − (inlier_number / N)^sample_size);

1.0 is not a principled value:

• Pixel estimators: SIFT at coarse scales (and ASIFT, dense matches) has inlier
  localization error of several pixels; inlier thresholds are commonly set to 3–5 px.
  With a 1.0-px reference, genuine inliers with 1–3 px residuals are excluded from
  the confident count, under-estimating the inlier ratio.
• Metric estimators (rigid transformation, plane, line): 1.0 is in data units,
  so identical geometry in mm vs m vs inches yields different iteration budgets.

The reference threshold affects the RANSAC search budget — the number of samples
drawn — and therefore, on hard problems, which model is found. It does not change
the per-hypothesis weight/loss math.

Previously only findEssentialMatrix_ touched it, and merely rescaled the arbitrary
1.0 default by the coordinate normalizer (still rooted in 1.0).

Fix

Derive the reference threshold from the noise scale, uniformly across all estimators:

magsac->setReferenceThreshold(EstimatorType::getSigmaQuantile() * sigma_max);

k = getSigmaQuantile() = sqrt(chi2_0.99(DoF)) makes this the statistical inlier
threshold τ = k·σ_max — scale-invariant and consistent with the MAGSAC++ weight/loss
cutoff (the weight is exactly 0 beyond k·σ_max). The normalized essential estimator
uses k · normalized_sigma_max.

Scope & relationship to other PRs

• Stacked on feat: add new model estimators (plane, PnP, radial homography) #45. Several affected estimators (PnP, radial homography,
  affine, planar/gravity essential) are added by feat: add new model estimators (plane, PnP, radial homography) #45. The plane/line reference-threshold
  fix already ships in feat: add new model estimators (plane, PnP, radial homography) #45 (new metric estimators, no existing behavior to regress). This
  PR covers the remaining pre-existing and pixel-domain estimators (rigid, fundamental,
  homography, PnP, radial, affine, essential variants). Rebase onto master after feat: add new model estimators (plane, PnP, radial homography) #45 merges
  for a clean diff.
• Behavior change → needs benchmarking. Unlike feat: add new model estimators (plane, PnP, radial homography) #45's plane/line additions, this PR
  changes the iteration budget of estimators that existing users rely on. Validate on the
  paper's datasets (Adelaide / EVD / homogr) for failure-rate / accuracy regressions before
  merging. A synthetic harness is available in the project's benchmarks/.
• Severity is low (search budget / runtime, not the per-hypothesis scoring), which is why
  it is separated from the bug-fix PRs.