add support for MultiFloats

[araujoms]

Tulip relies on IEEE 754 treatment of Infs, and therefore doesn't work with non-IEEE types like MultiFloats: https://github.com/dzhang314/MultiFloats.jl

For example, here Tulip relies on diving an Inf by a normal number remaining an Inf, but MultiFloats turn it into a NaN.

Tulip.jl/src/Presolve/Presolve.jl

Lines 282 to 283 in 59e77f0

	pb.lcon[i] /= rscale[i]
	pb.ucon[i] /= rscale[i]

Luckily the fix is simple: if we regard NaNs as Infs then everything works. I'd be happy to add tests if you're ok with this solution.

[mtanneau]

I remember having that exact issue and decided not to support MultiFloats. I can't remember exactly why though.
I believe it might be something about later comparisons with Inf, or something like a division by Inf that should have been zero.

Keeping Infs around was a fairly early design decision that sometimes comes back in moments like this.
Papilo keeps an additional vector of booleans that tracks whether a bound is finite or infinite, that might be a valid strategy as well.

[araujoms]

I see, we get a test failure in the unbounded example. It should be easy to fix though, one can simply do it manually if it's a single operation with Inf. Redesigning Tulip to use booleans would be a lot of work.

[araujoms]

I've tried a new approach now, defining some functions to do Inf arithmetic manually for non-IEEE floats. I think it's the optimal solution, as it doesn't affect currently working cases.

I just don't know whether I found all the places in the code where Inf arithmetic is needed. Tests pass, but I don't know how comprehensive they are.