Add the geometric term to the flex stretch stiffness

[smallquail]

K_stretch was the Gauss-Newton Hessian of the stretch force rather than its Jacobian. With elongation e_a = L_a^2 - L0_a^2 and force f = -sum_ab M_ab e_a grad(e_b)/2, the Hessian is

K = 2 sum_ab M_ab (s_a d_a)(s_b d_b)^T  +  sum_a Me_a (Laplacian_a (x) I3)

and only the first term was present. The second is proportional to the edge tension Me_a = sum_b M_ab e_b, so it vanishes at rest and grows with strain — the operator was first-order correct and no more.

Finite-differencing against -d(qfrc_passive)/dq on a mesh dilated by 5% gives an error of 7.8% of the force scale; with the geometric term it is exact to roundoff. The existing FlexStretchDerivatives did not catch this because it deforms at 5e-4 strain, where the two agree — its comments documented the omission as intended behaviour.

Only the tensile part is added. The geometric block is Me_a*[[I,-I],[-I,I]] over the edge's two vertices, which is positive semi-definite exactly when Me_a >= 0. A compressed edge would make K indefinite, and both consumers — the CG constraint solver and the PCG in mjd_effSolve — require an SPD operator. Because the sign condition is structural, the projection is a max(Me, 0) and needs no eigendecomposition. It is confined to the operator: mj_flexPassiveStretch keeps the full Me_a, so no force changes.

Both the matrix-free operator and the CSR assembly that the effective metric builds from are updated, since they must agree.

Impact

This changes how flexes with elastic2d="stretch" integrate under the implicit integrators and the effective metric. model/flex/bag.xml moves; model/flex/poncho.xml is bit-identical, because bending energy is quadratic in position and has no geometric term at all. Models with no flex elasticity are untouched.

The interpolated-flex path still uses its own Gauss-Newton approximation, which FlexInterpDerivativesDeformed asserts explicitly; that is left alone.

Testing

• New DerivativeTest.FlexStretchDerivativesTensile dilates the mesh so every edge is in tension (the clamp inactive, so the operator is exact) and finite-differences the operator against the force. It fails without this change at 7.02 vs a 0.009 tolerance.
• Full suite: 1342/1342 pass.