Search filtering

[irevoire]

To merge after #37

Related to #5

We could probably improve the search performance by a lot by storing in every split node all the IDs below it and choosing the branch accordingly. Tested and failed in #39

While filtering out half of the items:

Found 95832 vectors
Louis's query took 5.556166ms
Starts querying all documents ...
Making the stats
On average it took 3.420697ms to query a vector
The fastest query took 873.5µs
The slowest query took 18.778792ms
Standard deviation: 774.028µs

Louis's query got twice faster, and the fastest query also got way faster than before.
But the slowest and average are pretty much the same as before.

---

Without filtering anything (just to see the impact of deserializing roaring bitmap on every descendants):

Found 95832 vectors
Louis's query took 5.5425ms
Starts querying all documents ...
Making the stats
On average it took 3.693682ms to query a vector
The fastest query took 1.093916ms
The slowest query took 14.116625ms
Standard deviation: 691.014µs

Compared to main before this PR:

Found 95832 vectors
Louis's query took 5.133333ms
Starts querying all documents ...
Making the stats
On average it took 3.797529ms to query a vector
The fastest query took 1.048167ms
The slowest query took 26.041417ms
Standard deviation: 779.071µs

In term of database size we won ~10% on movies:

-rw-------@ 1 irevoire  staff   667M Dec  4 11:44 movies/data.mdb
-rw-------@ 1 irevoire  staff   604M Dec  5 17:54 movies_roaring/data.mdb

[Kerollmops]

bors merge 😂