Add tensor_to_scalar add_output overload (#142)

[petlenz]

Closes #142.

Scalar-from-tensor outputs (von Mises, dissipation, trace) were inexpressible: OutputDecl::expr held only scalar/tensor alternatives while the t2s emitter sat fully wired but unreachable from outputs. This adds add_output(name, expression_holder<tensor_to_scalar_expression>, role) producing a Kind::Scalar OutputDecl (renders as double &<name>_out through the existing ScalarOutput path — no canonical_arguments/render changes needed).

• Every OutputDecl::expr variant-visit site audited: SymbolValidationPass leaf sweep and the CodeEmitPass render arm gained t2s branches (LeafCollector/CodeEmitPipeline already had collect_t2s/.t2s()); NumSimMaterialTarget's two std::get sites now guard with holds_alternative and reject loudly ("does not support yet… add_output overload for tensor_to_scalar expressions (von Mises, dissipation) — scalar-from-tensor outputs are inexpressible #142") instead of bad_variant_access; AlgorithmicTangentPass is pre-guarded by its Kind::Tensor check.
• Validation on the new overload: duplicate names, role attributes, plus a rejection of roles carrying expected_rank != 0 (a t2s output is scalar-valued; roles::Dissipation passes, roles::Stress throws).
• Backends: standalone + MOOSE supported (MOOSE keys on Kind::Scalar → MaterialProperty<Real>); NumSimMaterialTarget rejects loudly pending its dual leaf-scoping/JSON work.

Tests (+7): signature/emission assertions, duplicate/role/leaf-validation negatives, NumSimMaterial rejection, and a compile-and-run gate (recipe T2sOutputCheck: tr_eps = trace(eps), vm_like = sqrt(w·(ε:ε))) numerically checked against hand-computed values.

Rebased onto current main. gcc-14 Debug (-Werror): ctest 328/328.