Optimize isqrt

[chipshort]

taken out of #2029
somewhat related to #1897

Benchmarked the gas costs and they look much better for big numbers, while only having a small regression for small numbers:

Uint64::isqrt	1.0	MAX	MAX / 2
old	4026600	4446000	4440900
new	4045800	4182450	4207950
rel.diff.	+0.5%	-5.9%	-5.2%

Uint128::isqrt	1.0	MAX	MAX / 2
old	4065600	24899400	24076200
new	4156200	4937850	6977250
rel.diff.	+2.2%	-80.2%	-71%

Uint256::isqrt	1.0	MAX	MAX / 2
old	4145700	114904200	114196500
new	4333500	7516500	14925450
rel.diff.	+4.5%	-93.5%	-86.9%

Uint512::isqrt	1.0	MAX	MAX / 2
old	4208550	567245850	565513200
new	4555800	15074250	37877400
rel.diff.	+8.2%	-97.3%	-93.3%

If we want to avoid the increase in the lower numbers really badly, we could try to figure out the points until which the old version is better and manually check if the number is below that threshold, but that's likely to change with different compiler versions, bnum updates, etc.

[aumetra]

LGTM in general, just this small nit

[webmaster128]

Super nice! Thanks for the benchmarks. Let's not complicate things for small savings for small numbers.

I can't follow the exact math right now but if this is well tested and the two of you agree, this is good to go.

[aumetra]

Looks good! Especially the solution with the Unsigned trait, very nice