Deprecate setrounding

[simonbyrne]

Fixes #26935.

This is the "minimal" change of deprecating Float32/Float64 support.

[JeffBezanson]

I'm not sure this is worth doing until we have a replacement.

[StefanKarpinski]

Except that it will be a breaking change, no? We don't want the existence of old, broken functionality to prevent us from introducing the new, correctly working functionality.

[JeffBezanson]

I thought we were just going to remove setrounding eventually, not reuse the name.

[JeffBezanson]

@simonbyrne Please advise --- can we keep this for now and add a totally separate better mechanism later, or is it necessary to remove this first?

[simonbyrne]

I'm not sure what the ultimate solution will look like: my best guess is something based on Cassette, similar to the Int64 -> Int32 literal conversion pass.

The main argument for removing it is that this design is fundamentally broken: basically it works for small toy examples, but if you try to use it with anything non-trivial the results aren't reliable.

[JeffBezanson]

Ok; if you think it's unreliable enough that we're better off removing it then we can do that.

[orkolorko]

Setting rounding modes for Float64 is now implemented through intrinsics in LLVM.

https://reviews.llvm.org/D74729

I think it may be worth to bring back this change to Julia base; directed rounding are used in some verified linear algebra algorithms to take advantage of LAPACK matrix multiplication to compute rigorous enclosures.

I wrote a small workaround package
https://github.com/orkolorko/SetRoundingLLVM.jl
but I think this should be brought back to Base.