- h5mread():
    * Support datasets with zero dimensions i.e. with ndim (a.k.a. rank
      in HDF5 terminology) = 0.
      system.file("extdata", "krumsiek11.h5ad", package="zellkonverter")
      contains several of such datasets.
    * Support compound datasets.

- Now that dimnames can be written to disk, make sure that matrix
  multiplication write the dimnames of the result to the HDF5 file.

- Keep investigating interaction between chunk and block geometries and their
  impact on block processing performance.

- Do some benchmarking (with special attention to memory usage).

- Add more unit tests.

