intel/fs: Frag coord lowering in NIR

We have or compute a UW gl_FragCoord.xy value, which seems suspiciously like what nir_lower_frag_coord_to_pixel_coord is doing for other HW. Using that pass, with a little rework of the system value representation, lets us reduce pixel shader payloads when only a subset of channels are used. While I'm here, move some gen4 interpolation specifics into NIR, producing better optimization.