test(e2e): split the tenant worker CPU time across QEMU threads - #3843

The node-join diagnostics phase samples the tenant worker's CPU
counters twice on failure, but it reads them from the compute
container's cgroup, and that cgroup holds every thread QEMU runs:
the guest's vCPUs, QEMU's main thread, and its IO and worker
threads. A container burning most of its quota while the guest
reports no progress is two incompatible findings wearing one
number, and no container-level counter separates them.

Add a collector that execs into each Running worker's compute
container, finds the QEMU process and copies every
/proc/<pid>/task/<tid>/stat line into the artifact verbatim, with a
legend beside them. Nothing is computed here. The legend names
where utime and stime sit and why counting fields over whitespace
reads the wrong column for exactly the threads the split is about:
a KVM vCPU thread is named "CPU 0/KVM", that space is inside the
comm parentheses, and every field after it shifts by one.

It runs inside the sampling loop the two existing subjects already
share, so it costs no second wait: both readings ride the interval
already being paid, and the rate comes from the two files plus the
stamps each one carries.

Every way the read can come up short is a named outcome rather than
an empty directory, and the legend and the pairing instruction are
withheld from each of them: telling a reader to subtract two files
asserts that both hold a whole reading. The probe shells out to
nothing but getconf, whose absence it names, because a probe that
needed cat would report a container missing coreutils as a
container whose QEMU had exited, and it counts what it managed to
read so that having read no process at all stays separable from
having read them all without finding QEMU. Its stderr is suppressed
on the walk rather than on either read: redirections are applied
left to right, so a 2>/dev/null after an input redirect is not in
place yet when opening the input fails, and a thread that exits
between the glob and the read would otherwise put a shell
complaint in the artifact beside a healthy capture. It also reports
a QEMU that was named and then gone before its task files could be
opened, because a process heading alone is enough to make the
capture non-empty and would otherwise carry a legend describing
columns the file does not have.

The Pod listing the cAdvisor captures already share now takes its
whole query as a parameter, so this collector reuses that read
rather than carrying a third copy of it, and asks for the Running
Pods only: a launcher left over from an earlier attempt costs the
sibling caller nothing and would cost this one a bounded exec and a
slot under its cap. The guard that holds the sampling group against
the room the tenant snapshot needs now counts every capped walk in
the loop rather than the two spelled the same way, so the new
subject is inside the arithmetic rather than beside it.

Assisted-By: Claude <noreply@anthropic.com>
Signed-off-by: Aleksei Sviridkin <3811295@gmail.com>