x86 backend pessimizes `_mm*_andnot*` usages

Example: https://godbolt.org/z/oa5qx864a

Reproduced here:

```c
void f(__m512i x, __m512i& a, __m512i& b) {
__m512i mask = _mm512_set1_epi8(0xf);
a = _mm512_and_si512(mask, x);
b = _mm512_andnot_si512(mask, x);
}
```

I'm trying to use `and` and `andnot` to save registers. But the compiler defeats any attempt I can think of to avoid this, it constant folds the "not" part of `andnot` and generates code that uses two different masks.

Run the linked Godbolt example and the C++ intrinsic reproducer first; no source file or test is named in the issue. Trace how the x86 backend handles _mm512_and_si512 and _mm512_andnot_si512, then compare the generated code with the reported two-mask behavior. Done means the reproduced case no longer pessimizes register use while preserving correct results.

Written by the indexing model from the issue text.