[X86][AVX512] missed fold scalar load > embed broadcast

https://godbolt.org/z/4hb7vfz5K

```c++
void foo(int32_t &a , int32_t &b, int32_t &c){
c = a > b ? a : b;
}

void foo2(uint32_t &a , uint32_t &b, uint32_t &c){
c = a < b ? a : b;
}

float foo3(int &a){
return a;
}
```
```asm
foo(int&, int&, int&):
vmovd xmm0, dword ptr [rsi]
vmovd xmm1, dword ptr [rdi]
vpmaxsd xmm0, xmm1, xmm0
vmovd dword ptr [rdx], xmm0
ret

foo2(unsigned int&, unsigned int&, unsigned int&):
vmovd xmm0, dword ptr [rsi]
vmovd xmm1, dword ptr [rdi]
vpminud xmm0, xmm1, xmm0
vmovd dword ptr [rdx], xmm0
ret

foo3(int&):
vcvtsi2ss xmm0, xmm15, dword ptr [rdi]
ret
```
should be
```c++
typedef int32_t i32 [[gnu::vector_size(16)]];
typedef uint32_t u32 [[gnu::vector_size(16)]];
typedef float f32 [[gnu::vector_size(16)]];

void bar(int32_t &a , int32_t &b, int32_t &c){
i32 x = i32{a,0,0,0};
i32 y = i32{b,b,b,b};
i32 z = x > y ? x : y;
c=z[0];
}

void bar2(uint32_t &a , uint32_t &b, uint32_t &c){
u32 x = u32{a,0,0,0};
u32 y = u32{b,b,b,b};
u32 z = x < y ? x : y;
c=z[0];
}

f32 bar3(int32_t &a){
return __builtin_convertvector(i32{a,a,a,a}, f32);
}
```
```asm
bar(int&, int&, int&):
vmovd xmm0, dword ptr [rdi]
vpmaxsd xmm0, xmm0, dword ptr [rsi]{1to4}
vmovd dword ptr [rdx], xmm0
ret

bar2(unsigned int&, unsigned int&, unsigned int&):
vmovd xmm0, dword ptr [rdi]
vpminud xmm0, xmm0, dword ptr [rsi]{1to4}
vmovd dword ptr [rdx], xmm0
ret

bar3(int&):
vcvtdq2ps xmm0, dword ptr [rdi]{1to4}
ret
```

In the first case (min/max/abs), this can save one load instruction.
In the second case (int2float), this can avoid the unnecessary second source operand.

Start with the Godbolt reproducer and compare the generated AVX-512 assembly for foo/bar, foo2/bar2, and foo3/bar3. Investigate the compiler's x86 vector optimization path for scalar-load folding and embedded broadcasts. Done means the first cases use a scalar load with an embedded broadcast, while the integer-to-float case avoids the unnecessary second source operand.

Written by the indexing model from the issue text.