[llvm] [NVPTX] Add IR pass for FMA transformation in the llc pipeline (PR #154735)

[Alex MacLean via llvm-commits]

https://github.com/AlexMaclean approved this pull request.

I'm still unconvinced by the stated rationale of vectorization here. It seems like we'd be able to achieve basically all real-world FMA vectorization opportunities by vectorizing fadd, fmul, and fma in SLPVecotrizer and then folding FMA during DAG ISel. 

However, I think doing FMA folding earlier seems nice for other reasons such as smaller IR and cross-block folding, so I'm in favor of this change. Please wait for additional non-NV review prior to landing. 

https://github.com/llvm/llvm-project/pull/154735