another bug affecting hdf5 common val counters

There is more corruption of Darshan's common value counters for the HDF5 module, this time only appearing in shared file benchmarks that require MPI reductions of these values. This appears to be less severe than #174 (closed) in that it doesn't come into affect until shutdown time (after other modules have already written their data to log) and it doesn't appear to affect counter values beyond the common access bins.