fix: use consistent Voigt order for the shell reference curvature

[oberbichler]

Problem

IgaShell3PAD builds the membrane metric (ref_a/act_a) and the strain transformation matrix T in Voigt order [11, 22, 12], and the actual curvature uses the same order:

act_b(act_a1_1.dot(act_a3), act_a2_2.dot(act_a3), act_a1_2.dot(act_a3));   // [11, 22, 12]

But the reference curvature built in add() used [11, 12, 22]:

ref_b(ref_a1_1.dot(ref_a3), ref_a1_2.dot(ref_a3), ref_a2_2.dot(ref_a3));   // [11, 12, 22]  <-- swapped

kap = T * (act_b - ref_b) applies the same [11, 22, 12] transform to both operands, so the mismatched ordering subtracts the b12 and b22 components across different slots — a wrong bending strain whenever the reference surface is curved (ref_b != 0).

Flat references (ref_b == 0) and the membrane element (IgaMembrane3PAD, whose ref_a/act_a are consistently [11, 22, 12]) are unaffected. This is why the existing baked reference test — a flat patch — does not change.

Fix

Reorder ref_b to [11, 22, 12] to match act_b and the transformation:

ref_b(ref_a1_1.dot(ref_a3), ref_a2_2.dot(ref_a3), ref_a1_2.dot(ref_a3));

Why the usual tests can't catch this

• The baked test_iga_shell_3p_ad asserts values produced by the element itself, and its reference is flat — the bug is invisible there.
• A finite-difference check of g vs f cannot catch it either: the AD differentiates whatever energy expression is written, so g stays consistent with f even when that expression is wrong.

Test

tests/test_iga_shell_3p_symmetry.py uses a value-level physical invariant on a curved reference: for an isotropic material the shell energy must be invariant under relabelling the parametric directions u1 <-> u2 (which only permutes the shape-function rows).

• Before the fix: f_orig = 0.4182 vs f_swapped = 0.3031 → fails.
• After the fix: equal → passes.

Cross-checked independently against a numpy reference implementation using the consistent [11, 22, 12] ordering (element f moves from 0.4182 (matches the buggy ordering) to 0.2237 (matches the reference)). Full suite: 44 passed, existing shell baseline unchanged.

Impact

Kirchhoff-Love shells with a curved reference geometry previously produced an incorrect energy, gradient and Hessian. Flat references and membranes were not affected.