fix: correct MAGSAC++ IRLS gamma weight (B2 + B3) in sigmaConsensusPlusPlus

[f-dy]

fix: correct MAGSAC++ IRLS gamma weight (B2 + B3) in sigmaConsensusPlusPlus

Problem

The IRLS refinement weight in MAGSAC::sigmaConsensusPlusPlus (magsac.h)
evaluates the gamma at the wrong argument — the same bug class as in
graph-cut-ransac's getScore, and inconsistent with this file's own
getModelQualityPlusPlus (which is correct):

• B3 — wrong noise scale. The weight uses sigma_max = current_maximum_sigma
  (= maximum_threshold). The MAGSAC++ paper, and getModelQualityPlusPlus, use
  sigma_max = threshold / k (k = √χ²₀.₉₉(dof)).
• B2 — table/precision mismatch. The index is computed with the 1e4
  precision against the coarse stored_gamma_values table (step 1e-3),
  reading Γ(1.5, 10·t) — the gamma at 10× the intended argument.

B2 and B3 partially cancel (10/k² ≈ 0.75 for DoF=4), which is why results are
still reasonable — but the IRLS weight is not the paper's, and disagrees with the
marginal scoring in the same class.

Fix

Mirror the already-correct getModelQualityPlusPlus:

// B3: weight noise scale = threshold/k; collection cutoff stays at maximum_threshold
const double weight_sigma_max = current_maximum_sigma * threshold_to_sigma_multiplier;
const double squared_sigma_max_2 = weight_sigma_max * weight_sigma_max * 2.0;
const double one_over_sigma      = C_times_two_ad_dof / weight_sigma_max;
...
// B2: index the fine table (step 1e-4) matching the precision constant
size_t x = round(precision_of_stored_incomplete_gammas * squared_residual / squared_sigma_max_2);
if (stored_incomplete_gamma_number < x) x = stored_incomplete_gamma_number;
weight = one_over_sigma * (stored_complete_gamma_values[x] - gamma_k);

The inlier-collection cutoff (current_maximum_sigma > residual) is unchanged
— the paper truncates the weight at r ≤ k·σ_max = threshold. getModelQualityPlusPlus
is untouched (already correct).

⚠️ Validation / scope

This changes results for the existing (DoF=4) estimators (it removes the
10/k² distortion in the IRLS refinement). Validate on Adelaide / EVD / homogr
(median error and failure rate, before/after) before merging.

Companion graph-cut-ransac PR fixes the identical bug in the duplicated
getScore weight: danini/graph-cut-ransac#53 — released
together. Independent of the per-estimator-DoF PR #45, which is adapted onto this
corrected base separately.