ART: Rewrite ParallelGC run-test

To better test what ParallelGC is intended to do, write a version
that will more fairly allocate objects between the threads, and
will let each thread OOM only once. Use a barrier to wait for
completion of all threads, and force a final GC by allocating some
more objects.

Bug: 16406852
Change-Id: I019ddc08515b9610c18d55280cd46cbe3ae3e5ac