Hand-unroll the SIMD dot product loop

[blambov]

This removes some data dependencies and permit much greater internal parallelism for the code.

On my machine this reduces the single-threaded dot product time for 1024 floats by one third, from 111 ns to 74 ns.

[jbellis]

In Bench and the SimilarityBench it's mostly a wash for me, but I'm happy to commit based on your results.

Bench results (omitting warmup on colbert-1M). The difference in build and search times is within the normal run-to-run variance.

Before:

ada002-1M: 983064 base and 9753 query vectors created, dimensions 1536
Build (full res) M=32 ef=100 in 136.62s with avg degree 31.99 and 0.65 short edges
Wrote [INLINE_VECTORS] in 8.90s
Uncompressed vectors
Using CachingGraphIndex(graph=OnDiskGraphIndex(size=983064, entryPoint=91031, features=INLINE_VECTORS)):
 Query top 99/99 recall 0.7468 in 2.08s after 36,863,426 nodes visited

After:

ada002-1M: 983064 base and 9753 query vectors created, dimensions 1536
Build (full res) M=32 ef=100 in 137.31s with avg degree 31.99 and 0.65 short edges
Wrote [INLINE_VECTORS] in 8.45s
Uncompressed vectors
Using CachingGraphIndex(graph=OnDiskGraphIndex(size=983064, entryPoint=91031, features=INLINE_VECTORS)):
 Query top 99/99 recall 0.7475 in 2.05s after 36,608,860 nodes visite

SimilarityBench:

Before:

Iteration   1: 132534573.353 ops/s
Iteration   2: 131339572.110 ops/s
Iteration   3: 137510039.044 ops/s

Iteration   1: 50.220 ns/op
Iteration   2: 50.490 ns/op
Iteration   3: 50.324 ns/op

After:

Iteration   1: 140738259.475 ops/s
Iteration   2: 124628756.103 ops/s
Iteration   3: 126568242.550 ops/s

Iteration   1: 53.262 ns/op
Iteration   2: 53.074 ns/op
Iteration   3: 53.379 ns/op