Use apfloat for full-precision sqrt

[RalfJung]

With rust-lang/rust#121793 and rust-lang/rust#118217, most float functions are now documented to have varying precision, so using host floats for them in Miri is actually fine. The only exceptions are mul_add (tracked at #2995) and sqrt -- those guarantee an infinite-precision result, so we should ideally use softfloats. (That said, using host floats is not wrong here since the host functions are in fact guaranteeing the infinite-precision result we are supposed to be emulating. Still, it seems cleaner to use softfloats for everything where precision is guaranteed.)

This is blocked on rust-lang/rustc_apfloat#14.