[llvm] [SelectionDAG] Support sign tracking trough `{S|U}INT_TO_FP` (PR #82808)

[Nikita Popov via llvm-commits]

nikic wrote:

> Proofs: https://alive2.llvm.org/ce/z/RQYbdw

Shouldn't the proofs be more along the lines of this?
```llvm
define i1 @src_sitofp_i16(i16 %x) {
  %xf = sitofp i16 %x to half
  %bc = bitcast half %xf to i16
  %r = icmp slt i16 %bc, 0
  ret i1 %r
}

define i1 @tgt_sitofp_i16(i16 %x) {
  %r = icmp slt i16 %x, 0
  ret i1 %r
}
```

I assume computeKnownBits() on floats is still a bitwise interpretation, so it shouldn't use FP comparisons (not that it makes a difference here).

https://github.com/llvm/llvm-project/pull/82808