perf: adds (some) arm64 assembly for 4 and 6 words modulus

[gbotrel]

Description

This PR refactors the code generation in field/ to allow some space for the various scenarios (generating only amd64, amd64 and arm64, none), depending on modulus size.

It adds element.Mul and element.Butterfly in arm64 assembly.

Context is to make the library and its direct users more cost efficient on arm machines.

How has this been benchmarked?

cherry picked, running on c8g.16xlarge;
bls12-377/fr

ElementMul-64                       25.6ns ± 0%  23.5ns ± 0%   -8.31%
ElementButterfly-64                 9.79ns ± 0%  5.32ns ± 0%  -45.69%

bls12-377/fp

ElementMul-64                       52.8ns ± 0%  44.8ns ± 0%  -15.10%
ElementButterfly-64                 12.4ns ± 0%   6.6ns ± 0%  -46.72%

bls12-377/fr/fft

FFT/fft_2**18bits-64                                    2.93ms ± 0%  2.03ms ± 0%  -30.83%  (p=0.029 n=4+4)
FFT/fft_2**18bits_(coset)-64                            3.31ms ± 0%  2.39ms ± 0%  -27.83%  (p=0.029 n=4+4)

bls12-377/fr/sis

15-20% improvements

bls12-377/

10-15% improvements on most ecc point operations

bw6-761/fp

BenchmarkElementMul-16     250           182           -27.18%

[cor]

This is fantastic! Thanks