Wrong results in some FMA cases

[RalfJung]

Due to this being an exact port of the LLVM library, this crate is affected by llvm/llvm-project#63895: some FMA produce wrong results. This seems worth tracking on our side as well, if only to make it easier to find, given that "produces wrong result" is the worst kind of bug.

[beetrees]

I've posted a PR to fix this in LLVM at llvm/llvm-project#98721. I've also made a Rust port of the fix at https://github.com/beetrees/rustc_apfloat/tree/fix-mul-add.

[RalfJung]

llvm/llvm-project#63895 has been fixed upstream. Can we get that fix into the Rust port as well? :) Cc @eddyb @wesleywiser

[RalfJung]

@beetrees could you make a PR for your branch?