When test run is canceled via signal, we want to send SIGTERM to trun
and let it kill/wait the tested processes. However until now we were
sending SIGTERM to trun and immediately checking whether some spawned
tested processes remain without giving any time for trun to complete first.

Fix this by moving the code that detects/cleanups leaked processes to
trun itself, and in the main nxdtest driver to only send SIGTERM to trun
and let trun do the cleanup/leaked processes detection.

This fixes erroneous "leaked process detected" as noticed by @jerome in
the following example:

    (nexedi/nxdtest!16 (comment 150835))
    $ nxdtest
    ...
    >>> sleep
    $ sleep 10
    ^C# Interrupt
    # stopping due to cancel
    # leaked pid=188877 'sleep' ['sleep', '10']		<-- "leaked" is a bit misleading here
    error   sleep   1.030s  # 1t 1e 0f 0s
    # test run canceled
    # ran 1 test case:  1·error

After the patch the output in this case becomes:

    $ nxdtest
    ...
    >>> sleep
    $ sleep 10
    ^C# Interrupt
    ok      sleep   0.604s  # 1t 0e 0f 0s
    # test run canceled
    # ran 1 test case:  1·ok

There is no more "leaked pid=...." emitted, but the change from "error" to "ok" status is wrong.
We will fix it in the next patch.

/reviewed-by @jerome
/reviewed-on nexedi/nxdtest!16