Replace cxa_guard fences with acquire loads

This seemed to be the only place in bionic where a fence on a
performance-critical path could be easily replaced by a stronger
load/store order constraint. Do so.

On x86 this should generate the same code either way. Based on
microbenchmarks of the relevant ARM instructions, this is currently
performance-neutral in this kind of context. But in the future, the
newly generated acquire loads should give us a performance benefit.

Test: Booted AOSP
Change-Id: I7823e11d6ae4fd58e0425244c293262e2320fd81