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

[f-dy]

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

Problem

MAGSACScoringFunction::getScore (scoring_function.h) computes the per-point
MAGSAC++ weight at the wrong gamma argument, due to two compounding bugs:

• B3 — wrong noise scale. It uses sigma_max = increasedThreshold (the
  inlier threshold itself). The MAGSAC++ paper — and this repo's own
  getModelQualityPlusPlus — use sigma_max = threshold / k, where
  k = √χ²₀.₉₉(dof). The function even defines threshold_to_sigma_multiplier = 1/k
  but never applies it.
• B2 — table/precision mismatch. The index is
  round(precision_of_stored_gammas · …) with precision_of_stored_gammas = 1e4,
  but stored_gamma_values is spaced at 1e-3 ([0, 36], 36 001 entries). So
  the lookup reads Γ(1.5, 10·t) — the gamma at 10× the intended argument.

B2 and B3 partially cancel (net factor 10/k² ≈ 0.75 for DoF=4), which is why
MAGSAC++ still works well in practice — but the weight is not the paper's.

Fix

// B3: weight noise scale = threshold/k; collection cutoff stays at threshold
const double sigma_max = increasedThreshold * threshold_to_sigma_multiplier;
const double squared_sigma_max_2 = sigma_max * sigma_max * 2.0;
const double one_over_sigma      = C_times_two_ad_dof / sigma_max;
...
// B2: index the fine table (step 1e-4) whose spacing matches the 1e4 precision
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);

• sigma_max = threshold/k drives both the gamma argument and the prefactor.
• The inlier-collection cutoff (residual > increasedThreshold) is unchanged
  — the paper truncates the weight at r ≤ k·σ_max = threshold.
• The lookup moves to the fine stored_complete_gamma_values table, whose
  1e-4 spacing matches precision_of_stored_incomplete_gammas.

After this, getScore's weight matches getModelQualityPlusPlus and the paper.

Tables: generated + trimmed to the useful interval

gamma_values.cpp is now produced by scripts/generate_gamma_values.py
(scipy.special), which emits the C++ with a header comment documenting how it was
generated and why its length. The two live DoF=4 tables
(stored_complete_gamma_values = Γ(1.5,·), stored_lower_incomplete_gamma_values
= γ(2.5,·)) are trimmed to the useful interval [0, k²/2] — 66 385 entries,
down from [0, 10] = 100 001 — because the weight is truncated at the inlier
cutoff r ≤ k·σ_max, so the argument never exceeds k²/2. The dead coarse
table (stored_gamma_values / precision_of_stored_gammas /
stored_gamma_number) is removed (it had no remaining consumer once getScore
moved to the fine table — verified across the whole gcransac repo).

The table is sized for the precise k = √χ²₀.₉₉(4) (the largest plausible
cutoff), so it works unchanged whether the consumer uses the imprecise k literal
or a precise one; lookups clamp to stored_incomplete_gamma_number.

Note on k precision (deferred to #50)

This PR keeps the existing imprecise k literal (k(DoF=4) = 3.64, with its
matching gamma_k) to stay focused on the B2/B3 weight bug. Precise per-DoF
k = √χ²₀.₉₉(DoF) is introduced by the per-estimator-DoF PR (#50). Because the
table is sized for the precise k, #50 can switch to precise k without
regenerating or resizing the table.

⚠️ Validation / scope

This changes results for the existing (DoF=4) estimators (it removes the
10/k² distortion). It must be validated on the paper's datasets
(Adelaide / EVD / homogr: median error and failure rate, before/after)
before merging.

Companion magsac PR fixes the identical bug in magsac.h::sigmaConsensusPlusPlus:
danini/magsac#49 — the two are released together. Independent of the per-estimator-DoF PR (#50),
which is adapted onto this corrected base separately. The now-unused coarse
stored_gamma_values / precision_of_stored_gammas can be retired once #50's
adaptation also stops using them.