Fix #2377: Reorganize the test helpers (where appropriate) to cuda_python_test_helpers

[mdboom]

This moves test helpers that are generically useful across the subprojects to cuda_python_test_helpers. It fixes some corner cases in the sys.path hack to import cuda_python_test_helpers. It doesn't install it as a proper package -- I had hoped that would be possible, but upstream pip does not support path dependencies in a dependency group. Using that would require pip to be improved or require everyone to use uv.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[mdboom]

/ok to test

[mdboom]

/ok to test

[mdboom]

/ok to test

[mdboom]

/ok to test

[juenglin]

LGTM

[github-actions]

Doc Preview CI
Preview removed because the pull request was closed or merged.

[rwgk]

My codex thinks this failure

https://github.com/NVIDIA/cuda-python/actions/runs/30987658226/job/92246084554?pr=2509

is due to this PR. See below. Is this already on your radar?

• Confirmed: this is unrelated to PR 2509 and is deterministic, not a flake.
- The job passes all 1,276 Pathfinder tests, then fails on the first bindings import with ModuleNotFoundError: No module named 'cuda'.
- Pixi never builds cuda-bindings, so PR 2509’s modified build hooks are never executed.
- The same failure already occurred in PR 2384, which was merged as 19e66499c2.
- PR 2384 changed both pixi.toml files without updating their lockfiles. With Pixi 0.73, the first automatic lock refresh produces an
environment missing the local cuda-bindings and cuda-pathfinder packages.
- I reproduced the exact failure from a clean checkout using the CI’s Pixi 0.73. A second invocation in the same checkout converges and
succeeds, but rerunning GitHub Actions starts fresh and repeats the first-pass failure.
- The new TOML table placement also appears to unintentionally move several existing test dependencies from Conda to PyPI.
I recommend fixing this separately from PR 2509:
1. Keep the existing test dependencies under [feature.test.dependencies].
2. Put only cuda-python-test-helpers under [feature.test.pypi-dependencies], assuming that dependency is intentional.
3. Regenerate both cuda_bindings/pixi.lock and cuda_core/pixi.lock with Pixi 0.73 until they are stable and pixi lock --check passes from
a clean checkout.
4. Exercise both source-build/import commands.
The same stale-lock problem exists in cuda_core, so fixing only the bindings lock would likely expose the core failure next. Optional
hardening would be to make CI use locked mode so future manifest/lock mismatches fail explicitly instead of silently rewriting the lock.

[juenglin]

@rwgk - Fix here: #2532