miri: implement most floating point math operations using soft floats

[eduardosm]

This is a preview of how it would look to implement math intrinsics using soft floats.

Deterministic non-determinism (this.machine.rng) is used to apply some random error to the results to model the non-guaranteed precision (except for sqrt).

cc #3898

[RalfJung]

Thanks for the PR! Implementation-wise it looks great (I would have some minor comments but it's not worth going into them yet).

As already mentioned, my main concern here is about pulling in a not-widely-used (and not-widely-vetted) third-party dependency to fix something that's not fundamentally wrong. The only part of this that is a bugfix is using a softfloat implementation for sqrt.

How large is an sqrt implementation built directly on rustc_apfloat, and could we reasonably carry it in-tree?

[bors]

☔ The latest upstream changes (presumably #4009) made this pull request unmergeable. Please resolve the merge conflicts.