hdf5 results: use attributes for small datasets

[jordens]

ARTIQ Feature Request

Problem this request addresses

A HDF5 dataset uses at least a few kB of disk space while most ARTIQ datasets in practice are smaller than the minimum HDF5 dataset overhead and many (e.g. scalars) are even smaller than their key. This is an inefficient use of disk space.

Describe the solution you'd like

• Investigate H5D_COMPACT layout.
• Investigate storing ARTIQ datasets as HDF5 attributes.
• Investigate limitations w.r.t. maximum object header size (64kB) and "dense attributes" as (implicit?) extension.
• Investigate when and how an ARTIQ dataset should be an HDF5 dataset and when an attribute. Merging of the key spaces.

Additional context

• 1272 for reorganizing hdf5 dataset groups

[airwoodix]

Could we also add compression to the discussion? As was pointed out on #m-labs, there's support for compression filters for hdf5 datasets. I did some tests to get an impression on readily available and working filters, as well as crude performance. It seems realistic to me to enable Zstandard compression by default on all array datasets (e.g. images), assuming that smaller data is stored as attributes. But this requires further testing, especially on Windows. More / better testing very welcome. At least exposing h5py create_dataset options would be great.