Add consistent GPU runner capability detection for kkt and occidata

[ocots]

The CI workflow already targets the occidata self-hosted GPU runner, but there is no shared GPU runner capability detection in the test suite.

If GPU/device-dependent tests are added or enabled, introduce a consistent capability predicate that recognizes both supported runners, kkt and occidata, for example:

get(ENV, "RUNNER_NAME", "") in ("kkt", "occidata")

The Handbook requires that a missing/non-functional GPU fails loudly on either runner:

if on_gpu_runner()
    Test.@test is_cuda_on()
end

The current ExaModels GPU tests use CUDA.functional() && ..., which silently skips the GPU assertion when CUDA is non-functional. Replace this with visible Test.@test_skip behavior on CPU/developer runners and a failing assertion on the expected GPU runners. Keep runner naming and related test documentation aligned with the CI workflow.

[ocots]

This issue is addressed by PR #326: #326

Implementing commit: 94351c5

test/runtests.jl now carries the single capability module described in the Handbook's philosophy/testing.md ("Capability-gated tests"), aligned with what CTSolvers already does: CUDA_FUNCTIONAL, ON_GPU_RUNNER and GPU_SOLVER_ARMED. ON_GPU_RUNNER matches the kkt / occidata substring of RUNNER_NAME, which the GitHub Actions runner agent sets by itself — no CI.yml or CTActions change was needed. The substring form is deliberate: our runners are registered as kkt-runner / occidata-runner, whereas the runs_on label in CI.yml is the bare kkt / occidata.

The two :exa GPU tiers previously sat behind a bare short-circuit on CUDA.functional(), so a correctly-skipped run on a CPU machine and a silently-broken one on a GPU runner produced exactly the same output: green, with the GPU tier simply absent. They now branch to Test.@test_skip and show up as Broken in the summary instead of vanishing.

A new test/test_environment_contract.jl enforces the contract centrally: the MadNLPGPU/CUDSS extension must be armed on every runner, a GPU device must be present on kkt and occidata, and the silent-guard anti-pattern must not reappear anywhere under test/. GPU_SOLVER_ARMED uses isdefined rather than CTSolvers' CUDSSSolver isa Type, because that symbol only exists once MadNLPGPUCUDAExt has loaded and the latter form would abort the whole run with an UndefVarError at module load instead of failing a single assertion.

Verified locally on CPU by faking the runner name, which is the only way to exercise the loud-failure path without a GPU: RUNNER_NAME=occidata-runner and RUNNER_NAME=kkt-runner both make the contract test fail as intended. The full suite is green at 2592/2592 with 9 visible skips, and the occidata GPU job runs on PR #326 through its label.

The fix is implemented and the test suite is passing. This issue will be closed automatically once PR #326 is merged.