Versioned-hdf5

Minor Changes

- Improve the performance of reading a reading a dataset that hasn't been
written to (e.g., reading from an already committed version).

- Fix Versioned HDF5 to work with h5py 3.3.

- Fix an issue that would occur when using np.datetime64 objects for
timestamps when the fractional part of the second was exactly 0.

Minor Changes

- Avoid some unnecessary precomputation in the hashtable object. This improves
the performance for files that have many versions in them.

Major Changes

- Support h5py 3. h5py 2.10 is also still officially supported.

- Add functionality to replay versions. This allows mutating old versions.

- Add new helper functions {any}`delete_version` and {any}`modify_metadata`
to delete a version and modify metadata on a dataset that must be the same
across versions, such as chunk size or dtype.

- Add helper function {any}`recreate_dataset` for more advanced version
replaying functionality.

Minor Changes

- Disallow accessing vfile[version] before version has been committed. Doing
so previously could lead to inconsistencies.

- Disallow accessing non-versioned groups from VersionedHDF5File.

- Better error messages from VersionedHDF5File when the underlying file is
closed.

- Remove codecov.

- Some improvements to the benchmarking suite.

Minor Changes

- Fix a bug where chunks could be deleted from a dataset.

- Workaround an upstream h5py bug in the tests.

Minor Changes

- Fix a bug where attrs could be deleted from a dataset.

Major Changes

- Many improvements to performance throughout the library, particularly for
datasets with many chunks, and for looking up versions by timestamp. This
also sets up potential future performance improvements by automatically
converting sparse staged datasets to fully in-memory.

- Add some additional benchmarks to the benchmark suite.