Use new ARM atomic opcodes where available, for DR and even for app

These are ideas from @algr :

ARMv8.1 adds atomic add opcodes with release-acquire semantics: LDADDA,
LDADDAL, LDADDL, and corresponding STADD*, etc. These are simpler to use
and preferable to the exclusive monitor loops in DR's own code. DR would
need to dynamically select whether these were available in the underlying
processor.

Furthermore, we could potentially translate some exclusive monitor loops in
the application to use these new atomic opcodes instead, avoiding the
exclusive monitor instrumentation issue #1698.