UnmountAllUnder reads all of /proc/self/mountinfo on every call, which contends with mounts on dense gVisor nodes #2950

On a node running a few hundred gVisor actors, ateworker-gvisor threads spend a lot of time blocked in the kernel while reading /proc/self/mountinfo (they show m_start in /proc/<tid>/wchan).

imagecache.UnmountAllUnder reads the whole mount table on every call, and the gVisor worker calls it several times per actor teardown: resetRunscStateAndPidFileDirs (for the runsc state dir and the pidfile dir) and the bundle dir cleanup in main.go. With about 1,000 mounts in the worker's mount namespace, each read seems to hold the mount namespace lock for a while and contend with the mounts and unmounts that other actors' restores and suspends are doing.

Measured on one c3-highcpu-192-metal node (5035a6fa, one gVisor worker, sweperf astropy-7336, closed-loop users), sampling blocked worker threads once per second:

Possible fix: find the mount points under the target directory with a walk that checks each entry with statx (STATX_ATTR_MOUNT_ROOT), without descending into mount points, and fall back to mountinfo if the walk is incomplete. In a prototype, the m_start waits went to 0 at 225 users and throughput improved about 6–8%. I'll send it as a PR referencing this issue.