vchoi_fork

[svn-r438] Changes for PHDF5 to support extendable datasets.

H5D.c:
    If a dataset, using chunk storage, is created/opened and if its data
    file is accessed via MPIO, allocate file space for the chunks immediately.
    Also do the allocation when a dataset is extended.  Must do it at these
    points because H5Dcreate, H5Dopen and H5Dextend are collective calls
    in MPIO access mode and are the only chance to get all MPI processes
    to synchonize their meta-data.
    Also disallow opening/creating datasets with compression option if
    MPIO is used.  Not easy to support concurrent compression writes in
    an efficient way.  But read-only access should be fine.  Will implement
    that later.

H5F.c:
    Disable file cache if MPIO is used by forcing cache size and number
    of elements to zero.

H5Fistore.c:
    Added two new routines, H5F_istore_get_addr and H5F_istore_allocate.
    H5F_istore_allocate supports allocating file space for the chunks
    of a dataset.  H5F_istore_get_addr returns the file space address
    of a chunk that has been allocated.
    Changed H5F_istore_read and H5F_istore_write to bypass the chunk
    cache read/write mechanism.  Instead, they do read/write directly
    with the file for only the elements requested.
    (Some debug statements, enclosed by #ifdef AKC, are still in the
    file.  I am still working on some code.  They would be cleaned
    out soon.)