[llvm] [NVPTX] Remove `NVPTX::IMAD` opcode, and rely on intruction selection only (PR #121724)

[via llvm-commits]

peterbell10 wrote:

Okay, I've reworked this to remove `NVPTX::IMAD` and instead rely on tablegen patterns to fuse the operations. I looked back in the history and it seems it was just always done that way since mad was added
https://github.com/llvm/llvm-project/commit/eafe26d0824c5282f8ff832f38e07e8c1eb92cf0#diff-cd5ad648f0fbe25520e97c78f6f302492b4d713effce854ec6393bb39bb58ded

Even today there are only one or two references to IMAD in combine patterns, so I think this is likely fine. The i128 lit test even seems to be doing a better job of generating mad now.

https://github.com/llvm/llvm-project/pull/121724