Switch to apfloat's FMA

[RalfJung]

Currently Miri uses host floats for FMA. However it would of course be better to use softfloats. :)

Unfortunately apfloat still has some bugs around FMA, see llvm/llvm-project#63895 / rust-lang/rustc_apfloat#11. Those probably need to be resolved and ported to rustc_apfloat before we should use it in Miri.

Cc @eddyb

[RalfJung]

Here's a testcase that currently behaves correctly in Miri but would fail if we were to switch to apfloat right now:

fn main() {
    let f = f32::mul_add(
        -0.000000000000000000000000000000000000014728589,
        0.0000037105144,
        0.000000000000000000000000000000000000000000055
    );
    assert_eq!(f, -0.0);
}

[eddyb]

Here's a testcase that currently behaves correctly in Miri but would fail if we were to switch to apfloat right now:

Notably, it also fails when building and running w/ rustc, if optimizations are enabled.