fix: use per-estimator degrees of freedom in MAGSAC scoring

[f-dy]

⚠️ Do not merge standalone — rescoped. The per-estimator DoF change is
statistically entangled with the B2/B3 IRLS gamma fix through the shared
k=√χ²₀.₉₉(n) (the B2×B3 10/k² cancellation is DoF-dependent), and shipping
DoF alone propagates B2 into the new DoFs. So this DoF work should land
together with the B2/B3 IRLS gamma fix (magsac
danini/magsac#49 + gcransac
#53) as one benchmarked PR. The unrelated
null-sample_ crash guard has been split out to
#52. This branch (fix/magsac-dof)
already carries DoF + B1 and is the basis for that combined PR.

Summary

The MAGSACScoringFunction in scoring_function.h hardcodes _DimensionNumber = 4 for all estimator types. This is only correct for the Sampson distance used by fundamental/essential matrix estimation (where the residual is derived from 4D data (x1,y1,x2,y2) and (r/σ)² ~ χ²(4)).

For other estimators, the degrees of freedom of the squared residual are different:

Estimator	Residual type	Correct DoF	Was using
Fundamental/Essential	Sampson distance	4	4 ✓
Homography	2D transfer error dx²+dy²	2	4 ✗
PnP	2D reprojection error du²+dv²	2	4 ✗
Radial homography	2D transfer error	2	4 ✗
Homography (affine)	2D transfer error	2	4 ✗
Rigid transformation	3D Euclidean dx²+dy²+dz²	3	4 ✗
Linear model (plane/line)	Scalar signed distance	1	4 ✗

The wrong DoF propagates to:

• k (the χ² quantile used as sigma-to-threshold multiplier): e.g., √χ²₀.₉₉(2) = 3.03 vs the hardcoded 3.64 = √χ²₀.₉₉(4)
• gamma_k (the upper incomplete gamma at the boundary)
• The entire lookup table (precomputed for Γ(1.5, x) = DoF=4 only)

The practical effect is that MAGSAC weights decay with the wrong tail shape — χ²(4) has a fatter tail than χ²(1) or χ²(2), so the scoring is more permissive of outliers than the statistical model warrants.

Fix

1. Add static constexpr size_t getDegreesOfFreedom() to each estimator, returning the correct value for its residual type.

2. Add precomputed E₁(x) table (gamma_values_dof1.cpp) for DoF=1, since E₁ has no closed-form expression in standard C++.

3. Use closed-form expressions for DoF=2 and DoF=3:

   • DoF=2: Γ(0.5, x) = √π · erfc(√x) via std::erfc
   • DoF=3: Γ(1, x) = e⁻ˣ via std::exp

4. Use existing stored_gamma_values[] table for DoF=4 (unchanged).

5. Dispatch via if constexpr in the scoring loop, with a static_assert in the else branch to catch unsupported DoF values at compile time.

Backward compatibility

• For DoF=4 (fundamental/essential), the code uses the exact same stored_gamma_values[] table — results are numerically identical.
• The gamma_values.cpp file is still included and used for DoF=4.
• No new dependencies; uses only <cmath> functions (erfc, exp, sqrt).
• Adding a model with DoF≥5 without providing a gamma implementation triggers a compile-time error.

Verification

chi2_0.99(1) = 6.6349, k = 2.5758  (was using k=3.64)
chi2_0.99(2) = 9.2103, k = 3.0349  (was using k=3.64)
chi2_0.99(3) = 11.345, k = 3.3682  (was using k=3.64)
chi2_0.99(4) = 13.277, k = 3.6437  (unchanged)

Split out: null-pointer crash in solver_linear_model.h

The null-sample_ crash guard that was originally bundled here has been
extracted to its own standalone PR, #52
(it has zero coupling to the gamma/DoF work and can merge independently).

DoF=1 weight singularity (fixed)

The E₁(x) table for DoF=1 has E₁(0)=∞ at index 0. When a small residual rounds to
index 0, the lookup returned ∞ → infinite weight. Fixed by capping table[0] = E₁(step)
(first finite entry) in gamma_values_dof1.cpp.

Known gamma-weight issues (NOT fixed here — flagged for maintainers)

Reviewing MAGSACScoringFunction in scoring_function.h against the MAGSAC++ paper
(Eq. 2.1) revealed two pre-existing inaccuracies, also present in magsac's IRLS:

• B2 — table-step/precision mismatch: the coarse table stores Γ(a, i/1000) (step 1/1000) but is indexed with precision_of_stored_gammas = 1e4 → the gamma is evaluated at 10× the intended argument.
• B3 — σ_max definition: the scoring uses σ_max = increasedThreshold directly; the paper uses σ_max = threshold / k.
• B2 and B3 partially cancel (10/k² ≈ 0.75 for n=4). Fixing B2 alone (precision = 1e3) is harmful — it removes the cancellation (argument becomes 1/k² of correct). Both must be fixed together. Not done here: changes results for all models and needs accuracy benchmarks.