feat: add metrics dot and cos

[cutecutecat]

Error bound Check

Check the probability that the inequality does not hold:
$est - err * 2.9 < real < est + err * 2.9$

Projection With Identity Matrix

vector + dim	metric	errors(attempt 1)	errors(attempt 2)	errors(attempt 3)
glove + 100	dot	19541 = 2.0%	29535 = 2.9%	19099 = 2.0%
glove + 100	cos	19656 = 2.0%	27026 = 2.7%	23352 = 2.3%
sift + 128	dot	100241 = 10.0%	104076 = 10.4%	97756 = 9.8%
sift + 128	cos	77400 = 7.7%	74132 = 7.4%	77400 = 7.8%
sift + 128	l2	104613 = 10.5%	88878 = 8.9%	135172 = 13.5%

Projection With Random Matrix

vector + dim	metric	errors(attempt 1)	errors(attempt 2)	errors(attempt 3)
glove + 100	dot	26702 = 2.7%	64816 = 6.5%	42559 = 4.3%
glove + 100	cos	25517 = 2.6%	11341 = 1.1%	54328 = 5.4%
sift + 128	dot	64626 = 6.5%	49139 = 4.9%	103020 = 10.3%
sift + 128	cos	88658 = 8.9%	92813 = 9.2%	89247 = 9.1%
sift + 128	l2	66545 = 6.7%	78286 = 7.8%	86253 = 8.6%

Benchmark

CPP

• RabitQ(sift+128+L2): QPS 1816 recall 0.9964

and

• RabitQ(gist+960+L2): QPS 414.02 recall 0.9789

Main

• CLI(sift+128+L2): QPS 1031 recall 0.9927
• PYO3(sift+128+L2): QPS 911 recall 0.9927

and

• CLI(gist+960+L2): QPS 361.20 recall 0.9763
• PYO3(sift+960+L2): QPS 323.15 recall 0.9763

With this PR

• CLI(sift+128+L2): QPS 1059 recall 0.9925
• CLI(sift+128+Dot): QPS 1024 recall 0.9479(L2 groundtruth)

and

• CLI(gist+960+L2): QPS 350.68 recall 0.9766
• CLI(gist+960+Dot): QPS 387.65 recall 0.0031(L2 groundtruth)

(Updated 9/13)With this PR

patch: no-residual

tag	top	QPS	recall
Glove-200-l2-main	100	274	0.9002
Glove-200-l2-main-patch	100	331	0.9010
Glove-200-l2-PR	100	284	0.9019
Glove-200-l2-PR-patch	100	344	0.9010
Glove-200-cos-PR	100	285	0.9014
Glove-200-cos-PR-patch	100	342	0.9012

[cutecutecat]

Implement

• precompute of centroids_square at kmeans_lookup_many_by_dot

[usamoi]

Please remove code that is used for dot distance + residual vector quantization.

[usamoi]

Why there is centroids_squares? If it's meaningless, remove it.

[usamoi]

Please resolve conflicts.