Use explicit schema versioning as decoding mismatched data can lead to aborts.

We just a single byte at the end as the dataset version which should suffice since only ever support two versions in each binary, meaning this is expected to roll over eventually.

Closes #111 (closed)