Faster EltwiseFMAModAVX512

[fboemer]

1.5x - 1.7x speedup on EltwiseFMAModAVX512IFMA using a few tricks also used in the NTT:

1. Switch to IFMA 52-bit mullo when computing vq_times_mod. The speedup is likely due to better pipelining on the same port

2. Use the negative modulus, neg_p. This helps compute q := a * b - q * p via two steps: tmp = a * b; q = fma(tmp, q, neg_p) rather than the previous three steps: tmp = a*b; tmp2 = q*p; q = tmp - tmp2 (see line 2 of Algorithm 4 in https://arxiv.org/pdf/2012.01968.pdf for background)

Additionally, for the case when the addition argument arg3 != nullptr, merging the two modular reductions led to a slight speedup.

On ICX with clang-12, I see

Benchmark	Before	After	Speedup
BM_EltwiseFMAModAVX512IFMA/1024/0	0.196us	0.116us	1.69x
BM_EltwiseFMAModAVX512IFMA/8192/0	1.53us	1.03us	1.48x
BM_EltwiseFMAModAVX512IFMA/16384/0	3.05us	2.06us	1.51x
BM_EltwiseFMAModAVX512IFMA/1024/1	0.268us	0.177us	1.51x
BM_EltwiseFMAModAVX512IFMA/8192/1	2.22us	1.47us	1.51x
BM_EltwiseFMAModAVX512IFMA/16384/1	4.43us	2.93us	1.51x

[hamishun]

LGTM