On termination of the worker pool in the main process, a SIGTERM is
sent from pool to worker. It was meant to terminate long-running
tests in the worker process. The signal handler on the worker side,
however, was only registered during test execution. During the
remaining logic (<1% of the time probably) the default system
behavior for SIGTERM would be used (which will likely just kill
the process). The ungracefully killed process might be killed while
writing to the results queue, which then remains with corrupted data.
Later when the main process cleans up the queue, it hangs.

We now register a default handler in the worker process that catches
the SIGTERM and also gracefully stops the processing loop. Like
that, the SIGTERM signal will always be handled in workers and never
fall back to SIGKILL.

However, a small time window exists when the SIGTERM was caught
right when starting a test process, but when the test-abort handler
was not registered yet. We keep fixing this as a TODO. Worst case,
the main process will block until the last test run is done.

Bug: v8:13113
Change-Id: Ib60f82c6a1569da042c9f44f7b516e2f40a46f93
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/3882972Reviewed-by: Alexander Schulze <alexschulze@chromium.org>
Commit-Queue: Michael Achenbach <machenbach@chromium.org>
Cr-Commit-Position: refs/heads/main@{#83101}