`reduce(min, Vec)` doesn't respect IEEE 754

[Moelf]

julia> a = SIMD.Vec((1.0, 2.0, NaN, 0.5))
<4 x Float64>[1.0, 2.0, NaN, 0.5]

julia> reduce(min, a)
0.5

julia> reduce(min, [1.0, 2.0, NaN, 0.5])
NaN

[KristofferC]

Everything in this package pretty much directly calls into LLVM so this should be reported there imo.

[Moelf]

depends on what we want though -- https://llvm.org/docs/LangRef.html#llvm-implementation they do have one minimum as oppose to minnum that has this semantics

[KristofferC]

I think https://llvm.org/docs/LangRef.html#llvm-vector-reduce-fmaximum-intrinsic is the relevant intrinsics which only seems available on julia nightly (LLVM 18).

#130