[benoitgoupilleau]

When a task exceeds its maxRunTime, the worker kills its processes straight away, so test harnesses can't log that they were aborted and their structured logs just stop.

This makes run-task abort the task itself shortly before the deadline:

• The task transform passes maxRunTime to the task in the TASKCLUSTER_MAX_RUN_TIME environment variable, for both docker-worker and generic-worker payloads.
• On POSIX, run-task schedules an abort at maxRunTime minus 5%, capped at 60 seconds early. It sends SIGTERM to its own process group: run-task is PID 1 under docker-worker/d2g and the group leader under generic-worker, which starts it with Setpgid.
• run-task ignores that SIGTERM itself, through a no-op handler rather than SIG_IGN so that its children still get the signal, and keeps forwarding the task's output while the harness shuts down.
• After an abort, run-task exits 1 whatever the command returned, whether the abort hit the checkout, the fetches or the task's command, and prints a line saying the task was aborted for exceeding its maxRunTime. A command that hasn't started by then isn't started, and a required command isn't retried.

The worker's own kill at maxRunTime is unchanged and remains the backstop if the task's processes ignore SIGTERM. Nothing changes on Windows, and tasks without TASKCLUSTER_MAX_RUN_TIME in their environment behave as before.

This is the upstream counterpart of the gecko change in https://bugzilla.mozilla.org/show_bug.cgi?id=2082559 (D338003), needed by the gecko tasks that run the upstream run-task as run-task-git.

Tests cover the margin calculation, invalid values, the SIGTERM handler, the process-group signal, a real child that receives SIGTERM while its output keeps flowing, the exit code after an abort, commands skipped after an abort, and the transform for both worker implementations.

[jcristau]

Requesting changes to take this out of the review queue, it needs an update to match the latest changes to the run-task-hg patch.

[benoitgoupilleau]

Updated to match the current D338003: run-task installs a no-op SIGTERM handler and always signals its own process group (no current_child fallback), run_command() returns 1 after an abort and doesn't start a command once the task is aborted, and retry_required_command() doesn't retry after one. Tests updated accordingly.