Add negative-path test coverage for intentional guard throws (#11)

[petlenz]

Fixes #11.

Every intentional error path could silently regress with no test failing. This adds tests/NegativePathTest.cpp, exercising the guards reachable through the public recipe / target API:

• SymbolValidationPass: output references an undeclared symbol; declared name is not a C++ identifier.
• MooseMaterialTarget: stateful input (roles::History); evolution equations without enable_local_newton; more than one consistent tangent; output named Jacobian_mult alongside a tangent; tensor rank ∉ {2, 4}.
• Duplicate add (symbol / output) and symbol↔output name clashes.

Each test asserts both the throw and a substring identifying the offending symbol/op, and was checked to pin the intended throw (not an incidental earlier one).

The originally-listed NUMSIM_CODEGEN_TENSOR_STUB / tensor_inner_product_to_scalar rows are documented as retired/unreachable (all Phase-A tensor stubs have landed; the surviving t2s stub has no recipe-level path — :/dot maps to the implemented tensor_dot). The stateful-output guard is unit-tested in its own PR (#15, MooseTargetTest.cpp), so this suite has no dependency on that change.