faster distance calculations via extrapolation

continue the work done with Eigen, but

• remove randomization (so, Eigen won't be needed)

• only use this optimization when dimensions already prepared for it
  • e.g randomized in the client (Matryoshka?)

• automatically detect if this is the case:
  • on inserts and on disk reads calculate the truncated and full distance, see if prediction works
  • may be only of the first X% of rows (when reading into the cache)