[llvm] [ARM] Allow FP reg conversion when copying Sx to Dx (PR #147559)

[Martin Storsjö via llvm-commits]

mstorsjo wrote:

This change seems to break cases that seem entirely legit (although slightly unusual) to me:
```c
#include <arm_neon.h>
void qt_convert_rgb888_to_rgb32_neon() {
  register uint8x8_t fullVector asm("d3");
  asm("" : : "w"(fullVector));
}
```
Here, `uint8x8_t` and the `w` constraint should be the right ones for the register `d3`.

https://github.com/llvm/llvm-project/pull/147559