It turns out that we had a bug where deleting an unpublished review

request would leave the DiffSet attached to the draft in the database,

but unreachable. These are now less unreachable, since FileAttachment

can link to the DiffSet. This was causing some crashes in the API

attempting to serialize a DiffSet that had neither a history nor a

draft.
This change makes it so we delete the DiffSet before deleting the draft.

Ran unit tests. The newly-added test failed before the change and

succeeds after.