[llvm] [NVPTX] support packed f32 instructions for sm_100+ (PR #126337)

[Alex MacLean via llvm-commits]

https://github.com/AlexMaclean commented:

> Rather than legalizing v2f32, we handle these four instructions ad hoc, so that codegen remains the same unless these instructions are present.

Supporting `v2f32` (similar to how we support `v2f16` for example) would be a cleaner and more extensible way to implement this change. 

https://github.com/llvm/llvm-project/pull/126337