job-exec: send SIGUSR1 to the IMP, not SIGKILL

Problem: RFC 15 states that the IMP handles SIGUSR1 by
sending SIGKILL to the entire cgroup.

For multi-user, send the IMP SIGUSR1 rather than SIGKILL after
shell signaling mechanisms have failed to clean up.

Update test faux imp shell script used in test.

src/modules/job-exec/job-exec.c

@@ -438,13 +438,21 @@ static void kill_shell_timer_cb (flux_reactor_t  *r,
 {
     struct jobinfo *job = arg;
     struct idset *active_ranks;
+    int actual_kill_signal = kill_signal;
+
+    /* RFC 15 states that the IMP handles SIGUSR1 by sending SIGKILL to
+     * the entire cgroup.  Sending SIGKILL to the IMP is not productive.
+     */
+    if (job->multiuser)
+        actual_kill_signal = SIGUSR1;
 
     flux_log (job->h,
               LOG_DEBUG,
-              "Sending %s to job shell for job %s",
-              sigutil_signame (kill_signal),
+              "Sending %s to %s for job %s",
+              sigutil_signame (actual_kill_signal),
+              job->multiuser ? "IMP" : "job shell",
               idf58 (job->id));
-    (*job->impl->kill) (job, kill_signal);
+    (*job->impl->kill) (job, actual_kill_signal);
     job->kill_shell_count++;
 
     /* Since we've transitioned to killing the shell directly, stop the

t/job-exec/imp-fail.sh

@@ -14,16 +14,6 @@ case "$cmd" in
         printf "test-imp: Going to fail on rank 1\n" >&2
         if test $(flux getattr rank) = 1; then exit 0; fi
         exec "$@" ;;
-    kill)
-        #  Note: kill must be implemented in test since job-exec
-        #  module will run `flux-imp kill PID`.
-        #
-        signal=$2;
-        pid=$3;
-        printf "test-imp: kill -$signal $pid\n" >&2
-        shift 3;
-        printf "test-imp: Kill pid $pid signal $signal\n" >&2
-        kill -$signal $pid ;;
     *)
         printf "test-imp: Fatal: Unknown cmd=$cmd\n" >&2; exit 1 ;;
 esac