Optimizations to KMeans and IVFBuilder

[hicder]

Summary

This adds 2 main implementation

• Parallelize posting list & kmeans builder, since they are independent and can be done in parallel
• SIMD the centroid computation

Result

Before this change, for query-optimized configs (max posting list size = 2000)

• Index time: 96s
• Query latency: 3ms

After this change, for the same config:

• Index time: 61s
• Query latency: 800 us (<-- no this is not a typo)