[llvm] [NVPTX] Add support for atomic add for f16 type (PR #84295)

[Artem Belevich via llvm-commits]

================
@@ -0,0 +1,17 @@
+; RUN: llc < %s -march=nvptx -mcpu=sm_70 | FileCheck %s
+; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 | FileCheck %s
----------------
Artem-B wrote:

Do these tests still work? If they do, then we have a bug somewhere because -mcpu=sm70 would set PTX to 6.0 and that should make `atom.add.noftz.f16` unavailable.

At the very least the test should have set ptx version via `-mattr=+ptx63`

Speaking of PTX version, you may want to have test cases for when the atomic instruction is not available. 

https://github.com/llvm/llvm-project/pull/84295