[AArch64] Failure to convert `x * y + z` into `fmadd` for `float16_t`

https://godbolt.org/z/voo9as94o
```c++
#include
#include

auto f16(float16_t x, float16_t y, float16_t z) -> float16_t {
// Intermediate results must be stored in variables to avoid conversions to
// float32_t
float16_t t = x * y;
return t + z;
}

auto f32(float32_t x, float32_t y, float32_t z) -> float32_t {
return x * y + z;
}

auto f64(float64_t x, float64_t y, float64_t z) -> float64_t {
return x * y + z;
}
```

```asm
f16(half, half, half, half):
fmul h0, h0, h1
fadd h0, h0, h2
ret

f32(float, float, float, float):
fmadd s0, s0, s1, s2
ret

f64(double, double, double, double):
fmadd d0, d0, d1, d2
ret
```

[Alive timed out](https://alive2.llvm.org/ce/z/2egtKJ), but since LLVM is happy to do the same transformation for float32 and float64, I assume it is valid.

Start by running the C++ reproducer in the linked Godbolt example and compare the AArch64 output for float16_t with float32_t and float64_t. Inspect the AArch64 compiler optimization path for floating-point multiply-add fusion; done means the float16_t sequence emits an fmadd instruction, subject to confirming legality because the linked Alive2 check timed out.

Written by the indexing model from the issue text.