3D beam torsion validation (circular section): mesh generation, SOFA vs FreeFEM vs Saint-Venant analytical comparison

[Fimache]

Summary

Adds a new validation case for pure torsion of a 3D beam with a circular
cross-section, cross-checking SOFA against FreeFEM and against the
Saint-Venant analytical solution.

Files added

• generate_beam3d_circular_tet.py — generates the P1 tetrahedral mesh
  (gmsh) of a circular cylinder, with tagged surfaces (Fixed, Loaded,
  Lateral) and volume (Beam).
• sofa_beam3d_torsion.py — SOFA scene: linear elasticity
  (LinearSmallStrainFEMForceField), one end fully clamped
  (FixedProjectiveConstraint), pure torque applied on the other end via
  consistent nodal forces derived from the shear traction field.
• freefem_beam3d_torsion.edp — same problem solved independently through
  FreeFEM's variational formulation of 3D linear elasticity, with the same
  traction distribution and a penalized clamp.
• compare_beam3d_torsion.py — runs both solvers, pairs nodes by
  coordinates, and compares displacement fields to each other and to the
  Saint-Venant analytical solution for torsion of a circular section.
• params.json / params_beam3d_torsion.json — geometry and material
  parameters.

Validation results

Comparison	Relative error
SOFA vs FreeFEM	1.76 %
SOFA vs Saint-Venant analytical	5.00 %
FreeFEM vs Saint-Venant analytical	6.67 %

• ux stays ~3 orders of magnitude below uy/uz, consistent with the
  theoretical absence of warping for a circular cross-section.
• uz grows linearly along the beam axis, matching the expected twist
  profile θ(x) = θ' · x.
• Boundary conditions are correctly enforced: displacement is exactly
  zero at the clamped end in both solvers.
• The gap vs. the analytical solution is attributed to (1) the tetrahedral
  mesh discretization of the circular section and (2) the boundary-layer
  effect near the rigid clamp, where Saint-Venant's free-torsion
  assumption doesn't strictly hold (Saint-Venant's principle).
• Small-strain validity was checked: total twist angle stays under the
  0.1 rad threshold for the parameters used here.