[AArch64] -ftrapping-math replaces vector fcmeq with scalar fcmp.

```c++
const float32x4_t zero = vdupq_n_f32(0.0f);
for (size_t i = 0; i < n; i += 4) {
const float32x4_t v = vld1q_f32(src + i);
const uint32x4_t nn = vceqq_f32(v, v);
vst1q_f32(dst + i, vbslq_f32(nn, v, zero));
}
```
With the -ftrapping-math flag enabled, the vector comparison is scalarized, even though it is written with intrinsics.
([-fno-trapping-math](https://godbolt.org/z/14oK9rchq) [-ftrapping-math](https://godbolt.org/z/8vG6Wfs86)).
This is presumably done to prevent flags from being lost during a floating-point exception. But fcmeq and fcmp raise the same flags (quiet equal).
The code proving this:
```c++
#pragma STDC FENV_ACCESS ON
...
uint32_t i = -1;
do {
++i;
float x;
std::memcpy(&x, &i, sizeof(float));
const float32x4_t xv = vdupq_n_f32(x);
uint32x4_t v_res;
uint32_t s_res;

std::feclearexcept(FE_ALL_EXCEPT);
asm volatile("fcmeq %[o].4s, %[a].4s, %[a].4s" : [o]"=w"(v_res) : [a]"w"(xv) : "memory");
const int v_flags = std::fetestexcept(FE_ALL_EXCEPT);

std::feclearexcept(FE_ALL_EXCEPT);
asm volatile("fcmp %s[a], %s[a]\n\tcsetm %w[o], eq" : [o]"=r"(s_res) : [a]"w"(x) : "cc", "memory");
const int s_flags = std::fetestexcept(FE_ALL_EXCEPT);

if (s_flags != v_flags || vgetq_lane_u32(v_res, 0) != s_res) {
std::cout << i << std::endl;
return 1;
}
} while (i != -1);
std::cout << "done" << std::endl;
```

(btw even -ffp-exception-behavior=maytrap removes vectorization)

clang 22.1.3 (homebrew) , apple m5

Reproduce the two linked Godbolt cases with clang 22.1.3 for AArch64, comparing the generated code with and without -ftrapping-math and -ffp-exception-behavior=maytrap. Trace the compiler's handling of the vceqq_f32 intrinsic and floating-point exception flags; done means the vector fcmeq is retained when its flags match scalar fcmp behavior.

Written by the indexing model from the issue text.