Micro optimization + const correctness

[rjanvier]

Hi. Thank you for the good work and making this Open Source!

This PR add some micro optimizations to the libs

• use std::array instead of std::vector to be sure to use stack allocated memory of the minimal size (3 floats)
• use reference to array instead of copys => On synthetic benchmark feature computation is between 5% and 8% faster. It won't be so much noticeable in real world scenario since NN query is the main bottleneck.

This address also minor style issues like const correctness, unused headers, use of std C++ arithemtic functions instead of C ones...