Avoid using leveldb values

[jacoscaz]

Appending size buffers at the end of each key rather than storing them as values might make for greater iteration speed assuming that accessing values is relatively costly compared to keys alone. Whether the ROI of doing this is high enough to justify the effort is TBD. This would also be a breaking change for existing datasets, thus candidate for a major release.

[jacoscaz]

closed via 5a39a65