Multi-threaded implementation of RMA for distributed shared memory


MPID_Win_fence(assert,win)

* Mark the window epoch as closed so that we can
  detect attempts to perform unsynchronized RMA operations.

  NOTE: In the case of multiple threads calling RMA operations,
  closing the window epochs without waiting for other threads to
  finish their RMA calls will likely result in a race condition if the
  application is not properly synchronized.  This is really an
  application error, so in the interest of high-performance, we make
  no attempt to minimize the non-determinism.

* Atomically move the remote handler call counters to a buffer and
  zero the counters.  This ensures that the 
------------------------------------------------------------------------