[llvm] [AArch64][SVE] Avoid transfer to GPRs for fp -> int -> fp conversions (PR #112564)

[Benjamin Maxwell via llvm-commits]

================
@@ -2421,6 +2421,42 @@ let Predicates = [HasSVEorSME] in {
   defm FSQRT_ZPmZ  : sve_fp_2op_p_zd_HSD<0b01101, "fsqrt",  AArch64fsqrt_mt>;
 } // End HasSVEorSME
 
+// Helper for creating scalar fp -> int -> fp conversions using SVE.
----------------
MacDue wrote:

I'll see about moving these patterns to C++, I'm not sure we'd want to lower `FP_TO_INT` and `INT_TO_FP`  separately though (vs a custom dag combine). If the user really did want the value in a GPR, I think we could end up turning. Though maybe we could just check the users while lowering the operation. 

```
fcvtzs w8, s0
// do something with w8
```

into:

```
ptrue p0.s, vl1
fcvtzs z0.s, p0/m, z0.s
fmov w8, s0
// do something with w8
```

https://github.com/llvm/llvm-project/pull/112564