First we send SIGTERM to leaked processes, and then, after timeout,
intend to send SIGKILL to the leaked processes and to the main test
process if it is still alive.

However there is a thiko in the code: we were sending SIGKILL only to
the main test process, not to all leaked ones.

-> Fix it.

The bug was there since 0ad45a9c (Detect if a test leaks processes and
terminate them).

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