[llvm] AMDGPU: Add round-to-odd rounding during f64 to bf16 conversion (PR #133995)

[Jay Foad via llvm-commits]

jayfoad wrote:

Should also delete the incorrect isel patterns in `VOP3Instructions.td`?
```
  def : GCNPat<(v2bf16 (bf16_fpround v2f64:$src)),
               (V_CVT_PK_BF16_F32_e64 0, (V_CVT_F32_F64_e64 0, (EXTRACT_SUBREG VReg_128:$src, sub0_sub1)),
                                      0, (V_CVT_F32_F64_e64 0, (EXTRACT_SUBREG VReg_128:$src, sub2_sub3)))>;
...
  def : GCNPat<(bf16 (bf16_fpround (f64 (VOP3Mods f64:$src0, i32:$src0_modifiers)))),
               (V_CVT_PK_BF16_F32_e64 0, (f32 (V_CVT_F32_F64_e64 $src0_modifiers, $src0)), 0, (f32 (IMPLICIT_DEF)))>;
```

https://github.com/llvm/llvm-project/pull/133995