Fix AVX2 distance function and build with AVX2 if available on host machine

[jparismorgan]

What

A few changes:

• Updates the build system so that if we have AVX2 support on the host machine, we will build with AVX2 support in the binary.
• Fixes several bugs in the AVX2 distance functions.

Based on:

• feat: [sc-52930] Set up vector search feedstock build variant with AVX2 #488
• Fix AVX2 distance function and build with AVX2 if available on host machine #499

Below are several sets of benchmarks that are run. You can also see more here: #499

The summary:

• VAMANA got much faster in both ingestion and query.
• IVF_FLAT had modest improvement.
• IVF_PQ doesn't seem to have a big improvement, and is worse on one benchmark, but not by much.

CI Benchmarks

C++ CI tests before (left) vs after (right). Note we go from 16 minutes to 10 minutes.

Python CI tests before (left) vs after (right). Note that while the cloud tests are not faster, the non-cloud tests are. i.e. test/test_distance_metrics.py::test_vamana_l2_distance goes from 41.01s to 23.19s. And test/test_ingestion.py::test_ivf_flat_ingestion_f32 goes from 28.90s to 21.48s.

Manual Benchmarks

Before:

After:

Before:

After:

Manual Benchmarks: VAMANA

Before:

After:

Before :

After:

Manual Benchmarks: IVF_FLAT

Before:

After:

Before:

After:

Manual Benchmarks: IVF_PQ

Before:

After:

Before :

After: