Simplifying downscaling logic again

This offers a small performance increase for all scaling factors while
providing the exact same quality at 2x (normal 4:2:0 use-case).

At 4x (when using a luma doubler) it is a bit worse but nothing
noticeable on real content. A nice side-effect is that the problem with
vanishing line-art seems to be a bit less severe now (I think it's due
to lower correlation, the older downscaling code was blurrier which
made the correlation between planes higher).

Again I may or may not revert this at some point, but I'm reasonably
happy with the change and it also reduces bloat a bit.

CfL_Prediction.glsl

@@ -25,47 +25,12 @@
 //!BIND HOOKED
 //!SAVE LUMA_LOWRES
 //!WIDTH CHROMA.w
-//!HEIGHT LUMA.h
-//!WHEN CHROMA.w LUMA.w <
-//!DESC Chroma From Luma Prediction (Downscaling Luma 1st Step)
-
-vec4 hook() {
-    float factor = ceil(LUMA_size.x / HOOKED_size.x);
-    int start = int(ceil(-factor / 2.0 - 0.5));
-    int end = int(floor(factor / 2.0 - 0.5));
-
-    float output_luma = 0.0;
-    int wt = 0;
-    for (int dx = start; dx <= end; dx++) {
-        output_luma += LUMA_texOff(vec2(dx + 0.5, 0.0)).x;
-        wt++;
-    }
-    vec4 output_pix = vec4(output_luma / float(wt), 0.0, 0.0, 1.0);
-    return output_pix;
-}
-
-//!HOOK CHROMA
-//!BIND LUMA_LOWRES
-//!BIND HOOKED
-//!SAVE LUMA_LOWRES
-//!WIDTH CHROMA.w
 //!HEIGHT CHROMA.h
 //!WHEN CHROMA.w LUMA.w <
-//!DESC Chroma From Luma Prediction (Downscaling Luma 2nd Step)
+//!DESC Chroma From Luma Prediction (Downscaling Luma)
 
 vec4 hook() {
-    float factor = ceil(LUMA_LOWRES_size.y / HOOKED_size.y);
-    int start = int(ceil(-factor / 2.0 - 0.5));
-    int end = int(floor(factor / 2.0 - 0.5));
-
-    float output_luma = 0.0;
-    int wt = 0;
-    for (int dy = start; dy <= end; dy++) {
-        output_luma += LUMA_LOWRES_texOff(vec2(0.0, dy + 0.5)).x;
-        wt++;
-    }
-    vec4 output_pix = vec4(output_luma / float(wt), 0.0, 0.0, 1.0);
-    return output_pix;
+    return LUMA_texOff(0.0);
 }
 
 //!HOOK CHROMA