Marking filters as optional

[nhz2]

The HDF5 format supports marking filters in the pipeline as used or unused for each chunk.

This might be useful, for example, if a chunk in an array is incompressible, it would be faster when reading to store the data uncompressed.

One option is to allow the user to pass a minimum compression ratio and size reduction required for a filter pipeline to be activated.

This could get even fancier; for example, we could allow passing multiple pipelines and some heuristics to pick the best pipeline per chunk.

[JonasIsensee]

This is true.
The most crucial component from the filter implementation -> return codes are already bubbled up to the caller.

However, optional filters are most relevant for chunked datasets.
So far, JLD2 only supports writing single-chunk datasets.
(I implemented reading of chunked datasets some time ago.)
Writing (multi-) chunked datasets is non-trivial because of the heap / table structures that need to be written to keep track of all the chunks.
This is hard - the hdf5 format spec does not give advice on good algorithms to construct these tables efficiently and how to edit them in case chunks are added incrementally.

For multi-chunk datasets, the filter pipeline is only written once for the dataset as a whole. The multi-pipeline would have to be merged and JLD2 would have to ensure that only the correct ones would get used.