[llvm] [IR] Store fast-math flags in Instruction (PR #188937)

[Joshua Cranmer via llvm-commits]

jcranmer-intel wrote:

> Though there are also plans to add more FMF, so it may not be a scalable solution.

I was exploring options to expand more FMF on the basis of wanting no less than 12 bits for flags, though I only have exploratory ideas for 3 more flags (no-overflow, optimize_libm, reassoc_abuse). I did find those bits in `Value` by rearranging it so that `SubclassOptionalData` stole the top bits of `NumUserOperands`, making it mostly a question of how few bits we need for user operands.

https://github.com/llvm/llvm-project/pull/188937