HDF5: Remove Support for Old Versions

[ax3l]

If we bump up our support to newer HDF5 versions, then we can optimize meta-data features & performance:

The call H5Pset_libver_bounds will set the guaranteed compatibility with older readers. The less old versions are supported, the more performant we can be.

Cool features in Quincy Koziol's ATPESC23 or HUG23 slides (to be linked):

• size of metadata structures H5Pset_istore_k, H5Pset_sym_k, etc.
• metadata aggregation in larger blocks H5Pset_meta_block_size
• Align objects in the file H5Pset_alignment
• Metadata cache size
• Page allocation and page buffering (see: File Space management)

We got to check, but can probably dump everything before 1.12 because our users & readers use recent pip/conda packages that mostly ship HDF5 1.14 already.

cc @jeanbez already increased quite a few of these buffers, but I was unaware of H5Pset_libver_bounds performance implications so far.

[jeanbez]

@ax3l let me know the minimum version you plan to support so I can open a PR with this. I can check the other optimizations since most have already been applied. I'm curious about their possible impact on openPMD.

[ax3l]

@jeanbez I think we want to drop at least everything older than HDF5 1.10.

I would like to bump to 1.12+ in the comings years, too, but some distros like Ubuntu is still a bit behind: https://repology.org/project/hdf5/versions

[jeanbez]

Got it, I believe that's a good way to go! I'll take a look into that and at the changes to include the H5Pset_libver_bounds