fix: derive MAGSAC reference threshold from sigma_max (existing estimators)

[f-dy]

fix: derive the MAGSAC reference threshold from sigma_max (existing estimators)

Problem

The pybind find*_ functions in src/pymagsac/src/magsac_python.cpp leave the MAGSAC
reference threshold (interrupting_threshold, magsac.h) at its constructor default
of 1.0. It 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 (homography, fundamental, essential): SIFT at coarse scales (and
  ASIFT, dense matches) has inlier localization error of several pixels; inlier
  thresholds are commonly 3–5 px. A 1.0-px reference under-counts genuine inliers.
• Metric estimators (rigid transformation, line): 1.0 is in data units, so
  identical geometry in mm vs m vs inches yields different iteration budgets.

This 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, merely rescaling
the arbitrary 1.0 by the coordinate normalizer (still rooted in 1.0).

Fix

Derive the reference threshold from the noise scale, uniformly:

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. Using getSigmaQuantile() (rather than a hardcoded k) keeps it correct before
and after any per-estimator DoF change. The normalized essential estimator uses
k · normalized_sigma_max.

Scope

Covers the pre-existing estimators on master: rigid transformation, fundamental,
essential, line, homography. Independent of other PRs — branched from and mergeable
into master on its own (22-line diff, one file).

The estimators added in #45 (plane, PnP, radial homography, affine,
planar/gravity essential) set this from the start in that PR, so they are not touched
here.

Note on validation

This changes the iteration budget of estimators that existing users rely on. It is
low-severity (search budget / runtime, not the per-hypothesis scoring), but should be
validated on the paper's datasets (Adelaide / EVD / homogr) for failure-rate / accuracy
regressions before merging.