[InstCombine] Incorrect clamp optimization propagates poison from `ninf` comparison

LLVM appears to miscompile this IR:

```llvm
define float @clamp_float_ordered_nonstrict_maxmin1(float %x) {
%cmp2 = fcmp nnan ninf olt float %x, 2.550000e+02
%min = select i1 %cmp2, float %x, float 2.550000e+02
%cmp1 = fcmp ole float %x, 1.000000e+00
%r = select nnan ninf i1 %cmp1, float 1.000000e+00, float %min
ret float %r
}
```
At`opt -O3`, it is optimized to:

```llvm
define float @clamp_float_ordered_nonstrict_maxmin1(float %x) {
%cmp2 = fcmp nnan ninf olt float %x, 2.550000e+02
%min = select i1 %cmp2, float %x, float 2.550000e+02
%.inv = fcmp oge float %min, 1.000000e+00
%r1 = select i1 %.inv, float %min, float 1.000000e+00
ret float %r1
}
```

The point is that the optimised op `%.inv` may lead to a `poison` result, then propagate to the return value, while the original will not.

Reproduce: https://alive2.llvm.org/ce/z/2uYLPY

Start by reproducing the transformation with opt -O3 and the linked Alive2 example. Then trace the InstCombine handling of the fcmp/select clamp pattern and add a regression test covering the ninf comparison; done means the optimization no longer propagates poison and the test preserves the original behavior.

Written by the indexing model from the issue text.