Parallelise writing of memh5 containers

[jrs65]

At the moment when distributed memh5 datasets are written to disk, they are done so serially, with each rank waiting for its turn to write. This is clearly a bit dumb when running on nice parallel filesystems like GPFS (nice might be a bit generous).

The way to work around this is to:

• Use one rank to create and pre-allocate the dataset as a contiguous dataset, also probably a good time to write out the attributes.
• Close the file.
• Redistribute the data to the slowest varying axis.
• Have each rank open the file, figure out the offset into the file for its chunk of data and then lock the range of the data it needs.
• Each rank writes its data and closes the file in parallel.