[petlenz]

Implements #216 (part of epic #221). No other PR has to merge first.

What and why

statev_map knew three history shapes: a scalar, a symmetric rank-2 tensor stored like a strain (plus stress_like from #91), and a minor-symmetric rank-4 tensor. Every tensor entry was rotated under DROT. Reference-configuration history (Cp, Fp, a fibre structure tensor), non-symmetric tensors, vectors and symmetric non-strains (conductivity, diffusivity) had no correct storage.

This PR generalises stress_like into a kind carried by the type:

• core/tensor_kind.h (new): enum class leg {spatial, material}, enum class voigt {none, engineering, plain} and kinded<T, V, Legs...>, a tmech::tensor<T, 3, sizeof...(Legs)> with the tensor's constructors, assignment and expression support. Aliases: stress_like, reference_strain, two_point, material_tensor, spatial_vector, material_vector, spatial_moduli, material_moduli.
• rotated(a, R) / rotate_legs<Legs...>(t, R): R acts on each spatial leg only, so R·v, R·t·Rᵀ, R·t, t·Rᵀ, or nothing.
• core/stress_like.h stays as a header that includes tensor_kind.h, so existing includes keep compiling.
• umat/statev_map.h: one entry builder (add_kinded<V, Legs...>) for every kinded type and for plain tensor<T,3,2>:
  • width 3 or 9 for voigt::none (all components, row-major), 6 otherwise (Abaqus order, shear doubled for engineering and as is for plain);
  • require_symmetric on pack for engineering/plain, no check for none;
  • a rotate step only if some leg is spatial, so has_rotatable_history() is false when only material-leg and scalar history exist, and the UMAT evaluator then skips the DROT step;
  • describe() marks each entry with its storage and legs: [engineering shear], [stress], [vector], [full], then [reference] (all legs material) or [two-point] (mixed legs).
  • The separate stress_like branch from STATEV: stress-like tensor history without the engineering-shear factor (#63) #91 is gone; stress_like goes through the same builder.
• umat/tensor_conversion.h: full_from_buffer / full_to_buffer (tmech full<3> adaptor) next to the strain/stress conversions.

A plain tensor<T,3,2> history keeps its meaning (engineering shear, spatial legs), and scalars and tensor<T,3,4> are unchanged. No existing material changes its STATEV layout.

Decisions

• stress_like was stress_like<T, Dim = 3>; it is now the alias stress_like<T> from the issue, because kinded is 3D only. Every user in the repository writes stress_like<value_type>.
• spatial_moduli and stress_like are the same type, as in the issue's sketch, so a spatial modulus is described as [stress] (stored without the shear factor).
• kinded rejects at compile time the combinations that cannot be stored: rank other than 1 or 2, a vector with six-slot storage, and six-slot storage of a two-point tensor (symmetry needs both legs alike).
• The full 9-slot layout is row-major {11 12 13 21 22 23 31 32 33}, the tmech full<3> adaptor.
• Rank-4 history keeps its own branch (all four legs spatial). No kind for it is needed yet.
• material_point / material_point_evaluator needed no change: they already ask has_rotatable_history() before the DROT step.

Tests

New tests/test_history_kinds.cpp (9 tests), with two test-only probe materials whose history outputs are kinded:

• stress_like<T> is kinded<T, plain, spatial, spatial>; rotated() acts on each spatial leg only (vector, symmetric, two-point either way round, material).
• With a 30° DROT, reference_strain, material_vector and material_tensor come back unchanged on both sides, while tensor2, spatial_vector, spatial_moduli rotate and two_point becomes R·F.
• two_point and material_tensor keep all 9 non-symmetric components through pack and unpack (raw slots checked); vectors round-trip through 3 slots.
• spatial_moduli stores D₁₂ while the tensor2 and reference_strain next to it store 2ε₁₂.
• reference_strain with a skew part is rejected on pack.
• Widths and describe() marks for every kind.
• has_rotatable_history() is false with only material-leg and scalar history.
• STATEV: stress-like tensor history without the engineering-shear factor (#63) #91's StatevMap.StressLikeHistoryKeepsItsShearComponentsForTheHost and all existing statev_map tests pass unchanged.

Gate

GCC 14 gate (with the ASan/UBSan test targets, -Werror): BUILD=0, 1458/1458 tests passed, 0 warnings.