llvm-reduce: Handle reducing FP values to nan

Prefer 0/1 over NaN, but it may make more sense to invert this as FP
operations with nan inputs can universally be folded into something
else.