Do not use atomic increment in allocation as fence.

A sequentially consistent fetch_and_add implemented with ARM v8 acquire
release operations is not a fence.  Don't use it as one.

The result may also be somewhat faster, since a sequentially consistent
increment requires more fencing than needed for the increment.

Bug: 16377103
Change-Id: I5b1add098d3488aa755f140612e54521b80aa749