Improve scheduler performance

Improve scheduler performance by flattening nonsensical topology layers
(layers with only one child don't serve any purpose).

This is especially relevant on non-AMD Zen systems after r322776.  On my
dual core Intel laptop, this brings the kern.sched.topology_spec table down
from three levels to two.

Submitted by:	jeff
Reviewed by:	attilio
Sponsored by:	Dell EMC Isilon

sys/kern/subr_smp.c

@@ -630,6 +630,15 @@ smp_topo(void)
 		panic("Built bad topology at %p.  CPU mask (%s) != (%s)",
 		    top, cpusetobj_strprint(cpusetbuf, &top->cg_mask),
 		    cpusetobj_strprint(cpusetbuf2, &all_cpus));
+
+	/*
+	 * Collapse nonsense levels that may be created out of convenience by
+	 * the MD layers.  They cause extra work in the search functions.
+	 */
+	while (top->cg_children == 1) {
+		top = &top->cg_child[0];
+		top->cg_parent = NULL;
+	}
 	return (top);
 }