MPI RMA implementation (device independent)

------------------------------------------------------------------------

Data structures

* MPI_Win

  An integer handle referring to a 'struct MPIR_Win'.

  The handle is converted to a MPIR_Win struct using a page table
  structure.

  Q: Should any of the bits in the handle be reserved for other
  purposes?

* struct MPIR_Win

  * comm - the communicator associated with this window.
  
    Q: Should this be a (MPI_Comm) or a (MPIR_Comm *)?
  
  * bases - array of bases addresses for the local windows (one per
    process)
  
  * sizes - array of local window sizes (one per process)
  
  * displs - array of displacement units (one per process)

  The following elements of the structure are extracted from other
  (more complicated) structures for the purpose of improving access
  time.

  * rank
  
  * np
  
  * base
  
  * size
  
  * displ
  
------------------------------------------------------------------------

MPI_Win_create

  
* Allocate a MPIR_Win structure and fill in the device independent portion
  (MPIR_Win).

  * Create a copy of the communcator provided by the user so the
    window has its own communcation context.
  
  Q: Which layer (MPI/ADI3) should manage (be responsible for
  creating, destroying, etc.) the MPIR_Win object?
  
  Q: Assuming the MPI layer managed the MPIR_Win object, should
  MPIR_Win be allocated using an MPID function?  Would this buy us
  anything?

* Acquire information about all of the processes involved in this window.

  * Get the window base address, size, and displacement unit from each
    process

  * If one more of those arrays contains the same value for all
    processes, then set the appropriate MPID_WIN_CONST flag.  When one
    of these flags is set, the local value can be used instead of
    performing a lookup in the array.

  NOTE: Acquiring the local window sizes and displacement units is not
  strictly necessary for all implementations, but it allows some of the
  error checking to be performed at the origin process.

* Let the device finish the creation of the window

------------------------------------------------------------------------

MPI_Win_free

* Q: do we need a reference count associated with the window object?
  If so, is it an error if the count is not one (1) when
  MPI_Win_free() is called?  Also, which layer is responsible for
  functions that manipulate the count and any associated mutexes?
  
* Give the device a chance to cleanup any device specific resources
  associated with the window

* Free the various data structures and Destroy the communicator
  associated with this window

------------------------------------------------------------------------