fix: moving the preserve mtiem test logic to Go for portability

[jpinkul]

The original bash version of the preserve mtime test doesn't work for Mac / Windows. For better portability I changed this to a go_test.

Changes are visible to end-users: no

Test plan

I've run this locally on a Linux machine to verify the basic functionality and ensured the test still fails if the preserve_mtime attributes are commented out. CI will be needed to test this on Mac / Windows.

[jpinkul]

I think this feature is behaving unexpectedly with Bazel's caching but I'm not sure why. This sequence is not working as I expected:

1. Run bazel clean --expunge
2. Comment out preserve_mtime = True from one of the inputs to the test.
3. Run the test with bazel test :test_preserve_mtime and it fails as expected.
4. Add preserve_mtime = True back and run the test again. It now passes as expected.
5. Comment out preserve_mtime = True a 2nd time and run the test again. This time the test still passes with the cached test result.

For that last test run I would expect the test to not be cached since a command line argument changed for the copy directory action this test depends on.

[alexeagle]

@thesayyn observes that RBE doesn't support modification times at all. I don't think that's the cause of the red CI here though. Sahin do you have any ideas?

Also, our CI here will trigger macos if macos is in the branch name and windows if windows is in the branch name. I expect that if you re-open the PR from a branch using the naming convention you'll be able to check those tests.

[jpinkul]

@thesayyn observes that RBE doesn't support modification times at all. I don't think that's the cause of the red CI here though. Sahin do you have any ideas?

Yeah I'm realizing this as well and I think remote caching is also incompatible with this feature. If the artifacts are copied from a remote cache then that process is probably going to reset the modify times. I tried adding the no-remote tag to the test targets and now the Linux version of these tests appears to be working on CI.

For that unexpected sequence I posted earlier I'm guessing that both actions with preserve_mtime = True and preserve_mtime = False were hashing to the same cache value since the contents on disk was identical in both cases. I haven't dug that deep into how Bazel handles the various cache keys though so this is a guess. I added the external tag to the test only though and it now passes / fails as I expected.

These two caveats should definitely be added to the documentation.

[jpinkul]

I documented these failure modes and pushed a new PR to trigger the mac/windows CI #908