fix: guard LinearModelSolver minimal fast-path against a null sample

[f-dy]

fix: guard LinearModelSolver minimal fast-path against a null sample

Problem

LinearModelSolver<2> / <3> (estimators/solver_linear_model.h) has a
minimal-sample fast-path in estimateModel:

if constexpr (_DimensionNumber == 2)
    if (sampleNumber_ == sampleSize())
        return estimate2DLine(data_, sample_, ...);   // derefs sample_[0..1]

if constexpr (_DimensionNumber == 3)
    if (sampleNumber_ == sampleSize())
        return estimate3DPlane(data_, sample_, ...);  // derefs sample_[0..2]

estimate2DLine/estimate3DPlane index sample_[0..k-1]. But estimateModel
explicitly supports sample_ == nullptr (meaning "use the first
sampleNumber_ points") — the general path right below it does exactly that via
const int sampleIdx = sample_ == nullptr ? pointIdx : sample_[pointIdx];.

So when the non-minimal fit is called with sample_ == nullptr and
sampleNumber_ == sampleSize(), the fast-path dereferences a null pointer and
crashes. This happens in MAGSAC's IRLS (sigmaConsensusPlusPlus), which refits
on all collected inliers passing sample_ = nullptr; when the inlier count is
exactly sampleSize() (3 for a plane, 2 for a line) it hits the fast-path.

Fix

Only take the minimal fast-path when the sample is non-null; otherwise fall
through to the general, null-safe path:

if (sampleNumber_ == sampleSize() && sample_ != nullptr)

Two-character change at each of the two sites. No behavior change for non-null
samples (the general path produces the same minimal-sample result).

Reproduction

Calling the non-minimal path directly with a null sample:

LinearModelSolver<3> solver;                 // 3D plane, sampleSize() == 3
solver.estimateModel(data, /*sample=*/nullptr, /*sampleNumber=*/3, models, nullptr);

• Before: AddressSanitizer: SEGV on address 0x0 at
  estimate3DPlane (solver_linear_model.h).
• After: returns normally (general path), no crash.

Scope

Standalone crash fix, branched from master, one file, zero coupling to the
gamma/DoF/scoring work. (It was previously bundled into the per-estimator-DoF
PR #50; it is extracted here so it
can merge independently.)