[X86][AVX512] suboptimal scalar source vecinit codegen on avx512

found from https://github.com/gcc-mirror/gcc/commit/f0bd50e831792692ffcedf288b312167a9339260
https://godbolt.org/z/he11xG4z6
```c
typedef float v4sf __attribute__ ((__vector_size__ (16)));

v4sf fabab(float a, float b) { return (v4sf){a,b,a,b}; }

```
-O3 -march=x86-64-v3:
```asm
fabab:
vinsertps xmm0, xmm0, xmm1, 16
vmovddup xmm0, xmm0
ret
```
-O3 -march=x86-64-v3 -mavx512vl:
```asm
.LCPI0_0:
.long 0
.long 4
fabab:
vmovddup xmm2, qword ptr [rip + .LCPI0_0]
vpermt2ps xmm0, xmm2, xmm1
ret
```

Start by compiling the provided C reproducer with -O3 -march=x86-64-v3, then repeat with -mavx512vl and compare the Godbolt assembly. Trace the x86 AVX512 vector-initialization code generation responsible for the vmovddup/vpermt2ps sequence. Done means the AVX512 case produces code at least as efficient as the non-AVX512 case, with regression coverage for the reproducer.

Written by the indexing model from the issue text.