nir_scale_fdiv: Respect vector swizzles

This pass that I wrote to avoid producing denorms (which would get flushed to zero) in the middle of a lowered div (which might be rcp + mul) didn't respect vector swizzles, so if it was run before lowering alu to scalar, it'd produce incorrect results. This is needed for OpenCL precision requirements.

Simple enough to fix the pass, so let's make it more robust.