[WIP] Enable AVX2 instruction set, change dependencies to improve performance

[Alexsandruss]

[Work in progress]

Proposed changes:

• Enable AVX2 support for Linux and Mac OS X platforms.
• Change build and runtime dependencies to MKL and Intel OpenMP to improve performance.

[rom1504]

if doing that, including a benchmark in the CI to check performance would make sense (do you know if AVX2 and MKL actually improve perf? I observed the opposite for MKL)

[Alexsandruss]

I will add run of some faiss benchmarks later. Can you provide details of case where MKL didn't improve performance?

[rom1504]

https://gist.github.com/rom1504/5a6a733dfd4772e497ea159e0395d0f9 can be a good starting point for a benchmark Using ivf index with train reconstruct and search