[llvm] [CodeGen] computeRegisterProperties - fix `bf16` to `i16` bitcast quieting NaNs (PR #224867)

[Matt Arsenault via llvm-commits]

================
@@ -571,6 +571,12 @@ class LLVM_ABI TargetLoweringBase {
   // values are always passed and returned as f32.
   virtual bool useFPRegsForHalfType() const { return false; }
----------------
arsenm wrote:

Why does this still exist? I thought this was supposed to go away as part of removing the half-to-float promotion? The coercion of the FP type in ABI contexts should be managed by the CConv handler and not need this special hook? 

https://github.com/llvm/llvm-project/pull/224867