x86/filmgrain: simplify post-horizontal filter blending

This commit makes a handful of minor changes:

• in horizontal blending, use shufps or vpblendd. If we change fewer pixels than can be used as one source operand for the given instruction (8 or 4 bytes), we abuse 0,32 as a edge/cur pair weight, so that the resulting blended register contains an unmodified cur grain. This replaces more complicated vpblendw + vpblendd or pand/pandn/por blending combinations.
• for scaling LUTs, always use psrld instead of pand, since the latter requires a register.