examples: unify doc samples across bindings + fix 3 bugs

[Grantim]

Summary

• Unifies C / C# / Python doc samples to match the cpp canonical: same inputs, same algorithm parameters, same output filenames. C samples drop CLI-arg handling in favor of hardcoded filenames, matching cpp/py convention.
• Fixes 3 bugs that the audit surfaced along the way.
• Bug-free same-named samples now demonstrate the same thing in every binding, so a reader flipping between language tabs in the Doxygen page won't see different algorithms.

Bugs fixed

File	Bug
examples/c-examples/MeshICP.dox.c	refMop was constructed from meshFloating instead of meshReference, so ICP was aligning a mesh onto itself. meshReference was loaded but never used.
examples/cpp-examples/MeshLoadSave.dox.cpp	if ( loadRes.has_value() ) error-check condition was inverted — the success branch printed the error and exited 1.
examples/python-examples/GlobalRegistration.dox.py	p2pl_inaccuracy = icp.getMeanSqDistToPlane(p2pt_metric) — passed the point metric where the plane metric belongs.

Other notable changes

• C# class-name fix: Collision / CollisionPrecise / CollisionSelf renamed to *Example so the Program.cs reflection dispatcher can actually discover them (it filters on EndsWith("Example")).
• Pedagogical completeness: MeshFixDegeneracies.dox.cpp and Triangulation.dox.cpp now save their results.
• MeshModification.dox.py now builds a torus programmatically (matching cpp) instead of loading mesh.stl; the parametrize entry in test_python_samples.py drops input_files accordingly.
• C# MeshDecimate.dox.cs switched to cpp canonical params (maxDeletedFaces=1000, maxError=0.05f, subdivideParts=64) plus the packOptimally() pre-step.

Known test breakage — requires S3 fixture refresh

The C# MeshDecimate test will fail against the current S3 autotest fixture (s3://data-autotest/test_data_2025-05-13/doc_samples/c-sharp/MeshDecimateExample/decimated_mesh.stl) because the canonical params produce a different mesh. The fixture needs to be regenerated with the updated sample and uploaded to a fresh dated snapshot, then the default URL in .github/actions/python-regression-tests/action.yml bumped. Will be done manually before merge.

Scope boundaries

This PR only touches changes that are fixture-independent (except for the one intentional break above). The following items are deferred to a follow-up because they would change existing fixture comparisons:

• C# MeshBoolean / MeshFillHole / MeshFixDegeneracies / MeshOffset logic alignment
• Python filename renames: LaplacianExample → LaplacianDeformation, NoiseDenoiseExample → NoiseDenoise, Triangulation_v3 → Triangulation
• Python output filename renames (e.g., decimatedMesh.stl → decimated_mesh.stl)
• New parametrize entries for Collision / SignedDistance / ContourTriangulation / MeshFixDegeneracies / MeshFromText / Triangulation_v2 in Python, and for Collision* / FreeFormDeformation / MeshICP / Triangulation in C#
• Adding the 14 missing C# .dox.cs samples to close parity with C++

Test plan

• Python regression tests (Linux / Windows / macOS) — expect MeshModification still green; MeshICP output unchanged (only getStatusInfo rename); GlobalRegistration output unchanged (only print-order + metric fix in stderr)
• C# samples test (Windows) — expect all current entries green except MeshDecimate which fails until S3 fixture refresh
• C++ / C examples build cleanly on all platforms (only source change; they aren't runtime-tested)
• Doxygen build has no broken \include warnings

🤖 Generated with Claude Code

[oitel]

The change drops code examples that might be useful for users. Maybe better to update other languages' examples instead?

[Grantim]

this sample seems much more complicated than other langs, do we actually need it all? mb just add outNewFaces usage to all langs?

[oitel]

Why is the code point preferred over an explicit symbol?