build: clean the warning set and enforce it in CI (#76)

[petlenz]

Closes #76.

The library warnings

-Wreorder	linear_elasticity.h: m_eps_name was declared after m_K/m_G but initialized before them, and m_eps is wired from it. Members initialize in declaration order regardless of how the initializer list reads, so this worked only by luck of m_eps being last.
-Wshadow x2	material_base.h: two constructor parameters named after the type aliases they shadow.
unused typedef	json_material_factory.h

Plus [[nodiscard]] on local_newton::solve and vector_newton::converged (backward_euler already had it) -- a dropped convergence flag is the bug this codebase has had most often -- and the dead m_material_handler, confirmed unread across include, tests and examples. The constructor parameter stays, since object_store passes it.

Why SYSTEM includes were needed first

I added -Werror for the cleaned classes, then found every remaining warning came from tmech, dozens of -Wshadow among them. As written, the CI change would have failed on a dependency we do not control, on its first run.

tmech, nlohmann and Eigen are now SYSTEM includes. Their warnings were never ours to act on, and they were burying the three that were.

The gate found nine more, in our tests

Local aliases duplicating the file-scope one, an unused handle, and an unbraced if around EXPECT_NEAR -- which expands to an if/else, so it was a genuine dangling else.

The tensor2 in test_j2_plasticity is not redundant: the file-scope one is fixed at 3 while that one is parameterized on Dim. Renamed rather than removed.

One fix nearly introduced a bug

This is the argument for the gate. test_statev_map.cpp had:

tmech::tensor<T, 3, 4> isotropic4(T K, T G) {   // shadows file-scope K, G
  C = 3.0 * K * IIvol + 2.0 * G * (IIsym - IIvol);

Renaming the parameters to bulk/shear left the body reading K and G, which silently rebound to the file-scope constants 166.67/76.92 -- the function would have ignored its arguments entirely. Caught on inspection and repointed.

The only caller passes exactly those values, so no test outcome would have changed either way. That is luck, not correctness, and it is precisely what -Wshadow is for.

CI

-Wall -Wextra -Wshadow -Wreorder, with -Werror for the three cleaned classes only. Deliberately not a blanket -Werror, which would fail on any new warning class a future compiler adds.

300/300 with -Werror active.

Note on overlap

This touches CMakeLists.txt (the target_include_directories lines) and #81 touches it too (the install section). They should merge cleanly; I will verify the merge result locally before either lands rather than assume it.