Optimize distance computations for IVFPQ

[NikolaosPapailiou]

Implemented a new query to quantized vectors distance implementation that is:

• Computing distance tables between each query and each pq_centroid
• Uses the query distance tables during query-vector distance computations

Tested the performance of all methods using SIFT-1M:

• asymmetric(previous default)
  
• symmetric:
  • ~1,3x speedup, accuracy reduction
    
• new distance based on query to pq_centroid distance tables:
  • more than 2x speedup, no accuracy reduction
    

When comparing latency for a larger dataset of 60M vectors of 200 dimensions, the original latency was ~100s and after this it is 9s.

[jparismorgan]

Looks great, thanks for adding the charts as well!

Could you also add a TODO to query_finite_ram() to update to this? Or you could update query_finite_ram() in this PR if you'd like, though it's fine with me to do in a follow-up once we support it via Python (I can do it then if you'd prefer, no worries).