Examples: Test cuda-core examples by calling main()

[mdboom]

The cuda_core examples are currently running in CI by exec'ing code. Now that they have all been refactored to have a main() function in #1674, we should be able to do that instead, as this comment suggests:

# TODO: Refactor the examples to give them a common callable `main()` to avoid needing to use exec here?

[mdboom]

Actually the problem is much worse: The current test as written is just an "import smoketest" of all the examples, but it doesn't actually run any of them. Adding the "bug" label here, because this is definitely a problem that our examples are being tested.

[leofang]

Actually the problem is much worse: The current test as written is just an "import smoketest" of all the examples, but it doesn't actually run any of them. Adding the "bug" label here, because this is definitely a problem that our examples are being tested.

Thanks for the good catch. We used exec for a purpose: We want to avoid the process setup/teardown overhead, which is very time-consuming for CUDA programs. Also, it allows us to save a bit of GPU memory when parallelizing tests with pytest-xdist. The exec usage has been used for several years, since the debut of cuQuantum Python and is now also used by nvmath-python.

IIUC what caused the samples to not run in the CI was due to the recent move to use __main__ (PR 1770). In this case, exec cannot execute the main() function.