Store the list of updated IDs directly in LMDB instead of a roaring bitmap

[irevoire]

Pull Request

After chatting with @ManyTheFish, we realized that deserializing a roaring bitmap, inserting an item, and serializing it again into LMDB every time we insert a new item wasn’t the smartest idea.
Ideally, we should write each document in a simple file or in RAM, but that adds a lot of complexity to the usage of arroy.
By simply writing the updated item ID directly in a database we were able to add 1M items in 1.7s instead of 18s+.

The drawback is that we use slightly more memory, but as a reminder, adding 150M vectors in one batch would only use around 0.5GiB (plus the LMDB internal structures).
This is clearly acceptable considering it’s memory mapped, and storing the vectors is going to take at least around 300GiB.

What does this PR do?

• Create a new kind of node ID updated where every associated ItemId represents the item ID of something we updated
• Update the rest of the code accordingly
• The tests still works 👌

[ManyTheFish]

Nice!