Change ANN library for a better one

[alexgarel]

The approximate nearest neighbour world is innovating a lot those years.

Annoy is far from being the best. See https://github.com/erikbern/ann-benchmarks

Change the ANN library for one of this.

Perf that interest us are on large dataset (we have more than 10M images) in cosine distance.

[raphael0202]

And with a high embedding dimension (~800 for CLIP models).