CI: investigate conda-cpp-tests jobs hitting the 40m timeout

Context

The `conda-cpp-tests` job in the `pr` workflow has been hitting the outer 40-minute `timeout` wrapper around `./ci/run_ctests.sh` (in `ci/test_cpp.sh`) and failing CI as a result.

Example failing run (PR #1035):

• Run: https://github.com/NVIDIA/cuopt/actions/runs/24721054667/job/72360181090?pr=1035
• Matrix: CUDA 13.1.1, Python 3.14, amd64, Ubuntu 24.04, H100, latest driver, latest deps
• Duration: 2026-04-21T16:55:08Z → 2026-04-21T17:40:09Z (~45 min wall-clock, killed by the outer `timeout 40m`)
• Conclusion: failure
• Runner: `609e-l-amd-g-h100-l-1-9rdw2-runner-v49qn`

As a short-term unblock, PR #1130 bumps the outer timeout from 40m to 50m. This issue tracks the underlying investigation — we should understand why the job is now running longer, not just raise the ceiling indefinitely.

Investigation items

• Break down the ctest run-time per test binary (e.g. `ctest --output-on-failure -V --timeout N` or parse the `Test.xml`) to identify which suites have regressed.
• Compare wall-clock of `conda-cpp-tests` on `main` over the last ~30 days — is this a recent regression or a gradual drift?
• Check whether slow tests are concentrated in one area (MILP, LP, routing GES, grpc integration) so we can target the fix.
• Confirm the regression is not specific to one runner pool or GPU (H100 vs. others in the matrix).
• Decide: (a) reduce test runtime at the source, (b) split into parallel jobs, (c) gate slow tests behind a nightly job, or (d) accept a permanently higher ceiling.

Acceptance criteria

• Root cause is documented on this issue.
• Either the offending tests are sped up / skipped on PRs, or a CI topology change is proposed and merged, such that the PR `conda-cpp-tests` job consistently finishes well under the outer timeout.

Related

• #1130 — bump the outer timeout from 40m to 50m as a stop-gap

Start with the pr workflow and ci/test_cpp.sh, then inspect how ./ci/run_ctests.sh is wrapped by the 40-minute timeout. Use ctest output or Test.xml to compare per-test durations, review conda-cpp-tests runs on main over the last 30 days, and compare runner/GPU results. Done means the root cause is documented and the job consistently finishes well under the outer timeout through a merged runtime or CI-topology change.