Speed up ivfflat build: use float instead of double for dot product

[pashkinelfe]

This is the first patch of patchset #178 as requested in the comments #178 (comment)

Regarding the conversion of the distance functions interface from double to float it will need modification of vector.sql interface which may need a major extension upgrade. If we're ready for this, I'd easily modify the patch as requested in #178 (comment) As of now I've left it as is for compatibility, and as this modifications will not make it much faster to build ivfflat index.

[pashkinelfe]

The results of tests are following:

1. ANN-benchmark: current master vs patch with float distance calculation (pgvector-float), vs patch with acos cubic Lagrange approximation with sign extension (pgvector-acos) vs both patches (pgvector-acos-float).

Base: dbpedia-openai-1M, index on inner product and queries using <#> operator, results dbpedia-openai-angular.png.

Float calculation has speedup in select queries probably because it is used calculation in query vector to ivfflat list vectors.
Acos approximation introduces slowdown probably because index quality regression at index build phase which is the only place where acos calculation is used.

2. Index build time (by stage): master vs patch with float distance calculation.

Speedup from patch for k-means and assign stage #1 is approximately equal (in percent)
For 1000 lists most of the time is spent at assign stage so, overall speedup from the patch is more due to assign stage.
For 4000 lists k-means and assign take almost equal time, so overall speedup from the patch is from k-means and assign staged equally.

3. Seq scan time master vs patch with float distance calculation and different distance functions.
   Using storage PLAIN.
   

[ankane]

Thanks @pashkinelfe, this is really great!! (both the performance win and the benchmarks) 🚀 🚀 🚀

Looks like it unlocks fused multiply-add on both aarch64 and x86_64 (additional benchmarks).

We should probably apply it to the cosine_distance function as well.

[ankane]

Updated cosine_distance in b710dc6.

[ankane]

Just fyi: switched to double for vector_norm in dab8f25 to avoid overflows. It did not affect performance in my testing (as it's only called once per sample and once per tuple during index builds).