Slightly expand the usage of VNNI256

[XZiar]

I am applying the changes on common patterns like _mm256_madd_epi16(_mm256_maddubs_epi16(a, b), ones) and _mm256_add_epi32(c, _mm256_madd_epi16(a, b)).

The original logics inside and outside HAVE_FANCY_SIMD are the same, so these changes should be safe (unless there's overflow handling outside of those regions...).

Actually there's already ultility functions like DotHelper::dot, using them could be neater...

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High

[ikawrakow]

Can you provide evidence that

1. The change does not break anything on an AVX2 system with VNNI
2. The change actually improves performance

As this change affects quite a few quantization types, I would appreciate if you tested all affected types.

Thanks.