Remove memory_barrier() calls from atomics.

This was unnecessary overhead to impose on all callers; the user
should handle these as needed on their own.

Also added some explanatory text to the documentation that highlights
that memory_barrier() is only needed across HW threads/cores, not
across program instances in a gang.

docs/ispc.rst

@@ -3880,6 +3880,11 @@ code.
 
     void memory_barrier();
 
+Note that this barrier is *not* needed for coordinating reads and writes
+among the program instances in a gang; it's only needed for coordinating
+between multiple hardware threads running on different cores.  See the
+section `Data Races Within a Gang`_ for the guarantees provided about
+memory read/write ordering across a gang.
 
 Prefetches
 ----------