Optimize _recreate_raw_dataset to use less memory; add versions method

[peytondmurray]

This PR optimizes a part of the _recreate_raw_dataset that currently can allocate significant memory for datasets with large numbers of versions. I also saw a large speedup in this section of the code post-refactor.

A VersionedHDF5File.versions method was added (a VersionedHDF5 should know what versions it has), which should return the same thing as versioned_hdf5.versions.all_versions but be much less verbose and more convenient.

Partially addresses #298. Deleting datasets can still be costly in terms of memory because the operation requires

1. Reconstructing the raw dataset, which entails keeping track of the raw chunks that need to be kept
2. Reconstructing the hashtable, which requires loading the inverse of the old hashtable as a new one gets built
3. Recreating the virtual datasets for the versions of the dataset we are keeping

[peytondmurray]

This change is low risk; merging now.