Fix GPU CI: propagate modeler backend option to test-problem builders

[ocots]

Summary

The self-hosted GPU runner (kkt) is now functional in CI, and it surfaces 13 test
errors across test_madncl_extension.jl and test_madnlp_extension.jl's GPU
testsets — testsets that were always silently skipped before (is_cuda_on() == false), so this was the first time they actually ran:

MethodError: no method matching MadNLPGPU.CUDSSSolver(::SparseMatrixCSC{Float64, Int32}; opt::MadNLPGPU.CudssSolverOptions)

Root cause: OptimizationProblem's Optimization.build_model methods in
test/problems/problems_definition.jl only forwarded modeler[:base_type] (Exa) or
nothing at all (ADNLP) to the wrapped build_*_model closures, silently dropping
every other modeler option — including :backend. So a GPU-strategy Exa
modeler still built a CPU-resident ExaModels.ExaModel, which then crashed when
handed to the GPU-only MadNLPGPU.CUDSSSolver.

Confirmed this is a regression (not a pre-existing test bug) against tag v0.4.17:
back then Modelers.Exa/Modelers.ADNLP were callable functors that forwarded
all options via Strategies.options_dict(modeler). Commit 5d01d1d
("test: update suite to the dispatch-based DOCP contract") replaced that pattern
with today's Optimization.build_model dispatch but dropped the general option
forwarding in the process.

Also checked CTDirect.jl's refactor/docp-dispatch branch for the same
regression in the real DOCP path: it already forwards options correctly
(Strategies.options_dict(modeler) / modeler[:backend]), so this fix is
confined to CTSolvers' own test harness — no CTDirect change needed.

Changes

• test/problems/problems_definition.jl: build_model for Modelers.ADNLP /
  Modelers.Exa now forwards the full Strategies.options_dict(modeler) to the
  builder closure instead of dropping options.
• test/suite/extensions/test_madnlp_extension.jl: the one remaining GPU testset
  using parameter-less (CPU-defaulting) Modelers.Exa(...) /
  Solvers.MadNLP(...) constructors now uses explicit {Strategies.GPU},
  matching every other GPU testset in the suite.
• test/suite/optimization/test_problem_definition.jl (new): contract-level
  regression test that catches this bug class on CPU-only CI, without needing
  real GPU hardware — it intercepts the builder closure's kwargs with fakes
  instead of building a real NLP model. Verified it fails against the pre-fix
  code and passes against the fix.

No src/ or ext/ changes — the strategy/backend metadata and defaulting logic
were already correct; only the test harness failed to use them.

Test plan

• Full CPU test suite green locally (2587/2587 passed), including the new
  regression test.
• New regression test manually verified to fail against the pre-fix code and
  pass against the fix.
• GPU CI job (test-gpu-kkt, gated by the run ci gpu label) — this is the
  real verification of the CUDSSSolver/GPU-array path; no CUDA hardware
  available locally.

🤖 Generated with Claude Code