Fix compute sanitizer in GitHub Actions CI

[bdice]

Description

@davidwendt noticed that the nightly builds were passing unexpectedly on GitHub Actions, after they had been regularly failing on a compute-sanitizer --tool memcheck step in gpuCI.

https://github.com/rapidsai/cudf/actions/runs/3880698576/jobs/6618987835#step:6:66543

After some investigation, I found the job was failing -- but for the wrong reason, and silently.

RAPIDS logger » [01/10/23 06:24:39]
┌──────────────────────────────────────────┐
|    Memcheck gtests with rmm_mode=cuda    |
└──────────────────────────────────────────┘

Running gtest ARROW_IO_SOURCE_TEST
ci/test_cpp.sh: line 95: compute-sanitizer: command not found
...

In this PR, I add the cuda-sanitizer-api package to our C++ test dependencies, which should fix the nightly tests (so they fail as expected). I also fixed a script bug that let the tests pass silently despite having failures.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[davidwendt]

Could we make this only run on one configuration? Say just one of the 11.8 ones perhaps?
conda-cpp-tests / tests (11.8.0, ubuntu20.04, arm64, 3.10, a100, 520)

[davidwendt]

According to the Jenkins logs for nightly runs, the compute-sanitizer step takes over 3 hours to run so it would be good to only run it for a single configuration and not all of them.

[codecov]

Codecov Report

❗ No coverage uploaded for pull request base (branch-23.04@8d8d0ee). Click here to learn what that means.
Patch has no changes to coverable lines.

Additional details and impacted files

@@               Coverage Diff               @@
##             branch-23.04   #12530   +/-   ##
===============================================
  Coverage                ?   85.65%           
===============================================
  Files                   ?      155           
  Lines                   ?    24810           
  Branches                ?        0           
===============================================
  Hits                    ?    21251           
  Misses                  ?     3559           
  Partials                ?        0           

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[bdice]

Thanks for that info @davidwendt, I didn't realize it took so long. Good idea to split it up and limit the configurations.

When I get a chance (may not be until next week) I will split the compute-sanitizer portion of the C++ test script into a separate file, add it to a separate workflow job using a custom script path (rather than having it be part of the C++ matrix tests), and have this launch only on one runner configuration, which we label gpu-latest-1. We also run Java tests and notebooks tests on gpu-latest-1 using a custom script, because they only need one configuration.

[wence-]

Just to note that I think we also want to, if we can, run with compute-sanitizer --tool initcheck (I think this is even slower than memcheck so running only on a single configuration is also necessary here)

[davidwendt]

Just to note that I think we also want to, if we can, run with compute-sanitizer --tool initcheck (I think this is even slower than memcheck so running only on a single configuration is also necessary here)

I would not recommend adding initcheck since I think it cannot reliably verify usage of uninitialized memory.
#12667 (comment)

[bdice]

Closing in favor of #12800 (already merged). Thanks @davidwendt for picking this up.