Optimize deletion by taking a transaction of the previous version of the database

Currently, the deletion in arroy runs in O(n) on the number of three nodes which is super slow.
The reason is that when we delete or replace an item, we cannot access the previous version of it, so we don't know where we'll find it, and we have to go through the whole tree.

We could maybe try to implement a slightly better version with https://github.com/meilisearch/arroy/issues/88, but it doesn't seem trivial and won't fix the issue for the updates of item.
Technically, even though it’ll be way faster, it's still O(n) on all the tree nodes, which won't scale forever.

The solution

By taking a read transaction of the previous version of the database, we would be able to read the previous version of the vector under the specified item ID and find it directly in the tree in O(log2(n)).

Start by tracing arroy’s deletion and replacement paths and the database transaction API; compare the current full-tree scan with issue #88. Define how a read transaction of the previous database version supplies the prior vector, then validate deletion and replacement behavior and the intended complexity.