Use more FMA in reciprocal iteration for precision

Description

Use more FMA in reciprocal iteration for precision

Typical Newton iteration for reciprocal is mul(x, fma(-a, x, 2)), which can also be written as fma(x, fma(-a, x, 1), x). Here two level of fma operations improves overall precision.

Reference issue

Additional information

Tested using https://gitlab.com/-/snippets/4903338, comparing the ratio of results 'nearer to' canonical 1/x result.

1 iteration (1000000000 random normal numbers and 10000000 denormal numbers):

(this means in the results of current method, 5.04% in normal numbers have better precision, while 20.73% with this change have better precision, and the rest, 74.23% are equal under both methods)

	Current Method	With This Change
Normal Numbers	5.04%	20.73%
Denormal Numbers	0%	0%

2 iterations:

	Current Method	With This Change
Normal Numbers	0%	21.32%
Denormal Numbers	0%	0%

1 iteration (using AVX512 rcp14 instruction):

	Current Method	With This Change
Normal Numbers	0.15%	21.42%
Denormal Numbers	0.11%	16.16%

2 iterations (using AVX512 rcp14 instruction):

	Current Method	With This Change
Normal Numbers	2e-07%	17.24%
Denormal Numbers	0%	12.96%