Orange Table-specific HDF5Reader

[stuart-cls]

Issue

Enable saving/loading the Orange Table data structure from the binary HDF5 container.
Based on the implementation used in the dask branch, but with the dask parts removed.

Related: #6356

Description of changes

• Add HDF5Reader with both read and write_file methods.
• Add round-trip test for HDF5Reader
• Add h5py to requirements
• Use existing .metadata sidecar file to store table.attributes if present.

Includes

• Code changes
• Tests
• Documentation

[stuart-cls]

I couldn't find a satisfactory solution to the .attributes problem (for the use case I care about) so I re-used the .metadata sidecar files for now, which is better than nothing.

[codecov]

Codecov Report

Attention: Patch coverage is 93.67089% with 5 lines in your changes are missing coverage. Please review.

Project coverage is 88.21%. Comparing base (5ada6c4) to head (63c71b3).
Report is 21 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #6791   +/-   ##
=======================================
  Coverage   88.20%   88.21%           
=======================================
  Files         327      327           
  Lines       71223    71301   +78     
=======================================
+ Hits        62825    62900   +75     
- Misses       8398     8401    +3     

[markotoplak]

Comments from @stuart-cls (from his email, just so that they do not get lost):

• Table.attributes -> I could try again to properly store this in the HDF5, instead of the .metadata sidecar. I got stuck trying to get the round-trip on a visible image to work :)
• Compatibility with dask branch: both opening files saved with that branch, and updating the branch to be compatible.

[markotoplak]

For domain_args we should certainly use the variable length h5py string data type. Probably for domain too.

[stuart-cls]

Done.

[stuart-cls]

Regarding the previous comments:

Table.attributes -> I could try again to properly store this in the HDF5, instead of the .metadata sidecar. I got stuck trying to get the round-trip on a visible image to work :)

In general this format isn't doing anything clever with nested dictionaries (see domain_args for example). It would be a lot of work to map this to HDF5, and this is the same problem with Table.attributes.

Compatibility with dask branch: both opening files saved with that branch, and updating the branch to be compatible.

I've tested both ways, it works fine. The new reader checks for "Orange" in the "creator" attribute, but falls back to checking that the "domain" group is there.