go from 623 seconds to 11 seconds on AMD64.
This is because we have no native atomic implementation on AMD64, 
so the rwlock-based implementation is used, which is  a bit slow.  
It will be probably be optimized for AMD64 quite soon, but for now 
this test case will do fewer iterations.


unittest/mysys/my_atomic-t.c:
  10k iterations instead of 1M no of iterations