What's being changed:

This PR introduces an ACORN1 like search.
Important features taken from ACORN are:

1. We totally ignore vectors not passing the filters to speed up search, since we do not calculate distance to them.
2. We use a multi hop approach to calculate neighbourhood of candidates to improve graph connectivity.

What we do not do from ACORN:

1. We do not require retraining.
2. We use random seeding for the entry points during search to speed up convergence to the filtered zone in the case of negatively correlated filters.

What to expect from the implementation:

1. No state on disk must be changed to support ACORN. You can simply upgrade from 1.26 to 1.27 and benefit from ACORN.
2. You don’t need any more resources than before, the vector index is the exact same as before.
3. Since it’s only a query time algorithm change, you can opt in or out at any moment.
4. For high recall settings we have seen almost 10x performance improvements for filtered queries for those that were very slow before. Still, you might expect 2 to 4x across most queries.
5. We have seen some small recall degradation under lab conditions. This might be negligible in real life settings though.

Finally, we add some tests with the super beir data. In the chart, Main stands for the old logic, while ACORN stands for the new logic, and the suffix indicate whether we use positive or negative correlation for the filter.

Review checklist

• Documentation has been updated, if necessary. Link to changed documentation:
• Chaos pipeline run or not necessary. Link to pipeline:
• All new code is covered by tests where it is reasonable.
• Performance tests have been run or not necessary.