Just to mimic this for experimental purposes, you could use external layout (H5Pset_external) and order the external references accordingly in one or more external files. Such a dataset wouldn’t be extendible (because of contiguous layout), but sufficient for testing. Then what? What’s the use case?
What I am exploring is if a HDF5 chunked file could be used as a Zarr shard and if there exist any limitations. For volumetric data, it is often useful to read nearby chunks via a contiguous read operation or range query. If the chunks could be arranged according to some kind of space filling curve, then a HDF5 file could be used in certain optimized read access patterns.