Correctly handle deletion of empty datasets

[peytondmurray]

This PR modifies how empty datasets are deleted, closing #244.

Changes

• As I was debugging, I broke out function definitions from inside delete_versions and _recreate_raw_data to help separate logic a little more. I don't think we gain anything by redefining _walk, _get_np_fillvalue, and _delete_dataset during each call, but I can change this back if it's just confusing.
• Added a bunch of type annotations for the functions I touched, it really helped with debugging once I knew what types different objects were.
• When the user deletes one or more versions, and those versions contain one or more datasets which don't exist in the versions being kept (and therefore the raw data needs to be deleted), we now skip recreating the raw data, recalculating hash tables, and recreating the virtual datasets associated with the dataset.
  This is done because empty datasets cannot be virtual; see 85507e9. In other words, empty datasets at _version_data/versions/<version>/<name> are real, and since they don't refer to data in _version_data/<name>/raw_data as they would if they held actual data, they can safely just be deleted. As far as I can tell, this is a limitation of h5py and hdf5 being unable to store empty virtual datasets.
• Added a test to cover the reproducer from Error when combining delete_versions and h5repack on empty Dataset #244.

[asmeurer]

It's been a while, but it seems correct. If you want to be extra sure you could try testing more extensive combinations of different versions including or not including the dataset with and without non-fill data.

I do rather wish I had made use of hypothesis when developing this library (#239), as its state machine strategies would be a perfect fit for extensively testing this sort of thing. It's something I would heavily suggest adding if we ever decide to add any significant new functionality.

[peytondmurray]

I really like the idea of making hypothesis do the heavy lifting for the tests. I'm going to try to spend a little more time doing some general upkeep on this library, so there's a good chance I'll do this in the future.

[ArvidJB]

I defer to @asmeurer, looks good to me, though!

Your observation that empty datasets are "special" and cannot be virtual seems to explain what's going on here. I think whenever we ran into this problem it did involve versions which were empty datasets.