[llvm] [InstCombine] optimize powi(X, C1)/powi(X, C2) with Ofast (PR #69807)

[via llvm-commits]

vfdff wrote:

I find the issue for this transformation has nothing to do with fast math flags, because it keep the `fast` attribute  : https://godbolt.org/z/97d8qhr89
  - In fact, `pow(a, 2)` will be transform into `fmul a, a`, which is not a Intrinsic ,so `II = dyn_cast<IntrinsicInst>(Op1)` will return 0

https://github.com/llvm/llvm-project/pull/69807