hdf5_der

[svn-r28042] Bug fix for failure in testpar/t_cache

Problem appears to have been caused by new code which 
dirties a cache entry unexpectedly at file close.  This
in turn throws off the write counts maintained by the 
test, causing it to fail.

Solution was to manually flush the cache prior to file
close, verify the expected writes, and then de-register
the callback used to obtain write counts from the cache.
After cache shut down, I added additional code to verify
that no test entries were written during the shutdown
process.

The net effect is to insulate the t_cache test code from
the shutdown process -- albeit at the cost of writing 
all test entries prior to the shutdown via a manual flush.  
As the code involved should be the same either way, this 
seems an acceptable solution.  However, we may want to 
re-visit it once file close proceedure has been tamed.

Tested parallel on mercury.

Note that a failure showed up the h5format_convert tests
on all builds after my first.  As all my changes were in
testpar, this almost certainly is un-related.  At a guess,
there is an issue with make distclean.  At Quincey's 
direction, I am proceeding with the checkin.

Note also that the following files

   tools/h5format_convert/testfiles/h5fc_edge.h5
   tools/h5format_convert/testfiles/h5fc_all.h5
   tools/h5format_convert/testfiles/h5fc_new.h5

were modified by my build, and are included in the 
subversion archive.  I reverted changes to them prior
to this checkin.