[llvm] [LLVM] Add `llvm.vector.reduce.fdot` intrinsic (PR #192174)

[Princeton Ferro via llvm-commits]

Prince781 wrote:

Looking at the test cases, it seems GPU codegen is better with `contract` and CPU codegen is better with `reassoc`.

https://github.com/llvm/llvm-project/pull/192174