[InstCombine] Wrong fold of `fabs` with `nsz`

_Opening a new issue for this because https://github.com/llvm/llvm-project/issues/214312 is owned by my NVIDIA account, which is temporarily deactivated._

`fabs nsz -0.0` is sometimes folded to `-0.0` which is no longer sound since https://github.com/llvm/llvm-project/pull/180906. This was fixed for `float`, `double`, and `half`, but still miscompiles with `fp128`, `bfloat`, `x86_fp80`, and `ppc_fp128`. Local Alive2 proof:
```
$ ./build/alive-tv -passes=instcombine temp/nsz.ll

----------------------------------------
define bfloat @f() {
#0:
%#1 = fabs nsz bfloat 0x8000
ret bfloat %#1
}
=>
define bfloat @f() {
#0:
ret bfloat 0x8000
}
Transformation doesn't verify!

ERROR: Value mismatch

NOTE: The counterexample is unique.

Example:

Source:
bfloat %#1 = #x0000 (+0.0)

Target:
Source value: #x0000 (+0.0)
Target value: #x8000 (-0.0)
```

@GabrielWu2004 is already working on this as a first PR. See https://github.com/llvm/llvm-project/pull/216232.

Start in the InstCombine area and reproduce the issue with the provided Alive2 command using temp/nsz.ll. Check the existing handling for float, double, and half against fp128, bfloat, x86_fp80, and ppc_fp128. Done means the listed cases no longer produce the unsound fold and the Alive2 transformation verifies.

Written by the indexing model from the issue text.