[AArch64] Update FP16 vector cmp costs

Without FP16, a fp16 v4f16 comparison will be converted to a v4f32 and back.
v8f16 get scalarized currently. Update the costs of v4f16 to match.