test: uniformize monorepo onto OrdinaryDiffEq canonical structure

[ChrisRackauckas-Claude]

Uniformizes the NonlinearSolve.jl monorepo onto the OrdinaryDiffEq.jl canonical structure: SafeTestsets group-dispatch testing (no ReTestItems), all test deps in [extras]/[targets].test (no test/Project.toml), and centralized CI workflows threaded with a repo-specific group env var.

Canonicalization done

Tests / framework (ReTestItems -> SafeTestsets)

• Removed all @testitem/@testsetup/@testsnippet/@run_package_tests/ReTestItems/TestItemRunner usage from the root and all 10 sublibraries.
• Each test item is now @safetestset "<title>" include("<file>__itemN.jl"). The item body runs at module top level (matching the original @testitem semantics) — this is required so BenchmarkTools @ballocated solve!($cache) interpolation resolves the package globals; an inline @safetestset begin ... end evaluates the interpolation in the wrong module and fails. Test bodies are preserved byte-for-byte, then Runic-formatted.
• @testsetup/@testsnippet modules are extracted to setup_<name>.jl and included by the items that referenced them via setup=[...].
• Every runtests.jl reads NONLINEARSOLVE_TEST_GROUP (falling back to GROUP) and group-guards its includes. Root test files carry per-item group guards so the mixed core/downstream/nopre/bounds/verbosity/cuda tagging is preserved exactly.
• Root test/runtests.jl is now a _detect_sublibrary_group dispatcher (mirrors OrdinaryDiffEq): a GROUP naming a lib/<X> activates that sublibrary, develops its in-repo [sources] transitively (incl. the umbrella root for leaf libs), and Pkg.tests it with the sublib group env set; otherwise it runs the root's own @safetestset groups. Heavy/optional group deps (ModelingToolkit, Enzyme/Mooncake/SciMLSensitivity, CUDA) remain Pkg.add-on-demand to keep the default matrix light.

Project.toml

• Dropped ReTestItems/TestItemRunner from every [compat]/[extras]/[targets].test; added SafeTestsets. No test/Project.toml anywhere (the test/trim/ trimming subproject is unrelated and untouched). No [compat] floors lowered.

test_groups.toml (added per sublibrary so the project-model matrix matches real groups)

• NonlinearSolveSciPy: only wrappers (it has no Core tests — critical, else the default Core job would run nothing).
• BracketingNonlinearSolve, SimpleNonlinearSolve: Core + adjoint.
• Others: Core on [lts,1.11,1,pre] (Aqua/ExplicitImports run inside Core; declaring real groups avoids a spurious empty QA job).

Workflows

• SublibraryCI.yml: group-env-name: NONLINEARSOLVE_TEST_GROUP, check-bounds: auto.
• DowngradeSublibraries.yml: group-env-name/group-env-value: Core (no allow-reresolve; auto-discovers lib/*; sibling+stdlib skip list kept).
• CI_NonlinearSolve.yml (root, GROUP-dispatched over the root's own groups): threaded group-env-name: NONLINEARSOLVE_TEST_GROUP, check-bounds: auto. No hardcoded sublibrary rows — sublibs are covered by SublibraryCI's project model.
• Downgrade.yml: already downgrade.yml@v1, group: core, no allow-reresolve — unchanged.
• GPU.yml: kept as-is. The SimpleNonlinearSolve CUDA path could move into the project model via a cuda test_groups.toml runner entry, but the root NonlinearSolve CUDA job is not a sublibrary, so both bespoke gpu-v100 jobs are kept. They set GROUP=cuda, which the new runtests honor via the GROUP fallback.

Root [sources] (the 4 "missing" sublibraries are intentionally NOT added)

• SciMLJacobianOperators is only a transitive dep (via NonlinearSolveFirstOrder), like RosenbrockTableaus in OrdinaryDiffEq, which is not hoisted into root [sources].
• SCCNonlinearSolve, NonlinearSolveHomotopyContinuation, NonlinearSolveSciPy are leaf libraries that depend back on the umbrella root (<root> = {path = "../.."}). Adding root->leaf edges would create cycles in the path graph (forbidden by the audit's rule B). All 10 sublibraries are still discovered and tested by SublibraryCI's lib/* scan.

Verified locally (isolated depot, Julia 1.11)

• TOML-parsed every changed Project.toml; confirmed all [targets].test deps appear in [extras]/[deps], zero test/Project.toml remain, zero @testitem/@testsnippet/ReTestItems remain. YAML-parsed all changed workflows; parsed all test_groups.toml; parse-checked all 217 converted .jl files. Runic-checked all 210 changed .jl (pass).
• Instantiated the root project and 3 sublibraries (NonlinearSolveBase, NonlinearSolveSpectralMethods, BracketingNonlinearSolve) — [sources] path graph resolves.
• Ran, via the new SafeTestsets dispatch with NONLINEARSOLVE_TEST_GROUP=Core:
  • NonlinearSolveBase Core: Testing NonlinearSolveBase tests passed (Aqua 11/11, Explicit Imports 3/3, Banded vcat 1/1, Termination 1/1, standardize_forwarddiff_tag 2/2, maybe_wrap_nonlinear_f 4/4, EnzymeExt 6/6 + 4/4, PolyAlgorithm 9/9, Bounds transform 8/8).
  • NonlinearSolveSpectralMethods Core (exercises @testsetup + @ballocated $cache): tests passed (DFSane 12/12, Iterator 2/2, NewtonRaphson Fails 2/2, Kwargs 3/3, Termination 27/27).
  • BracketingNonlinearSolve Core (exercises @testsnippet): tests passed (Interval Nonlinear Problems 17861/17861, Muller 13/13, Aqua 11/11, etc.).
• Ran the root's smallest group via the new dispatcher (GROUP=misc): Testing NonlinearSolve tests passed (Aqua 10/10, Explicit Imports 3/3) — confirms per-item group guards select only misc items.
• Unit-tested the root _detect_sublibrary_group/_find_lib routing (e.g. scimljacobianoperators -> sublib, simplenonlinearsolve_adjoint -> SimpleNonlinearSolve/adjoint, core/misc -> root-own).

What CI must confirm

• The full SublibraryCI project-model matrix across all 10 sublibraries x groups x Julia versions (lts/1.11/1/pre), including the heavier wrappers/SciPy/Homotopy/FirstOrder/QuasiNewton suites not run locally.
• The full root CI matrix (core/downstream/wrappers/misc/nopre/trim x lts/1.11/1 x ubuntu/macos), DowngradeSublibraries, and root Downgrade.
• The bespoke GPU.yml CUDA jobs on the self-hosted runner.

Ignore until reviewed by @ChrisRackauckas.