This commit implements Cgroup CPU research watcher for notifying CPU
throttling of cgroup. When configured, the adaptive calculator polls
this watcher periodically. When it returns a backoff event, the
calculator adjusts the concurrency limit accordingly.

Cgroup doesn't have an easy way to gauge how bad the CPU is percentage.
Hence, this watcher compares the recorded total throttled time between
two polls. If the throttled time exceeds 50% of the observation window,
it returns a backoff event. The watcher uses `throttled_time` (CgroupV1)
or `throttled_usec` (CgroupV2) stats from the cgroup manager.