By always enabling the debug defines in the test_itemrepository we
easily ran into all kinds of assertions. Some where wrong, but most
where at least partially correct and uncovered a wealth of connected
issues which I could not detangle from another. As such, this patch
sadly does a few things together, as otherwise the tests would always
fail. Thankfully, at the end, I managed to run not only the tests
with the debug assertions enabled, but could also parse a large
real-world application and save/load it from the cache! Previously,
both would always lead to assertions in various places.

Some of the issues that I uncovered:

- We never marked the tails of the monster buckets anywhere, meaning
  when we got an item request that fell into a bucket index that was
  merged with others to form a monster bucket we didn't detect it
  and silently corrupted the data by just writing the new item data
  into the bucket as-if it was an uninitializ...