[llvm] [LLVM][NVPTX]Add BF16 vector instruction and fix lowering rules (PR #69415)

[Han Shen via llvm-commits]

================
@@ -629,7 +657,7 @@ let hasSideEffects = false in {
                 ToType, ".bf16 \t$dst, $src;"), []>,
       Requires<!if(!eq(ToType, "f32"),
                    // bf16->f32 was introduced early.
-                   [hasPTX<71>, hasSM<80>],
+                   [hasPTX<70>, hasSM<80>],
----------------
shenh10 wrote:

However, the constraint here won't eliminate the generation of cvt.f32.bf16. Do you think we should instead constraint it in hasBF16Math constraint? Though it is not that rigorous but without bf16->f32 cast it's meaningless to support other instruction though.
```cpp
  bool hasBF16Math() const { return SmVersion >= 80 && PTXVersion >= 71; }
```



https://github.com/llvm/llvm-project/pull/69415