test/bench: osu-style bcast benchmark

The barrier often does not exit uniformly especially if node-topology is
in play. This affects different collectie algorithms differently, thus
using the combined latency doesn't hides too much details for algorithm
comparisons.

The osu microbenchmarks measures collective latency individually then
reduce for min, max, and average. Why it is still suceptible to barrier
behavior, it does provide more details for some insights comparing
different algorithms.

test/mpi/bench/bcast.def

@@ -5,6 +5,28 @@ include: macros/mtest.def
 page: bcast, bench_frame
     data: buf, size, MPI_CHAR
 
-    &call measure_with_barrier
-        MPI_Bcast($(data), 0, comm)
+    $global root=0
+    $(if:0)
+        &call measure_with_barrier
+            MPI_Bcast($(data), root, comm)
+    $(else)
+        $if grank == 0
+            $call header_coll_latency
+        &call foreach_size
+            $my tf_min, tf_max, tf_avg, tf_sigma
+            $(set:MIN_ITER=0.001/tf_max)
+            &call coll_warmup
+                measure_bcast(iter, root, comm, buf, size, &tf_min, &tf_max, &tf_avg, &tf_sigma)
+                tf_dur = tf_max
+            $if iter < 100
+                iter = 100
+            measure_bcast(iter, root, comm, buf, size, &tf_min, &tf_max, &tf_avg, &tf_sigma)
+            $if grank == 0
+                $call report_coll_latency, size
+
+fncode: measure_bcast(int iter, int root, comm, buf, size, pf_min, pf_max, pf_avg, pf_sigma)
+    &call measure_coll_latency, iter
+        MPI_Bcast($(data), root, comm)
+    $(for:min,max,avg,sigma)
+        *pf_$1 = tf_$1