CustomPixelShaderEffect: honor queried output sub-rect

Previous impl always returned m_lastOutputRect (the FULL output rect)
from MapOutputRectToInputRects regardless of which sub-rect D2D asked
for. That was the safe-with-TRIVIAL_SAMPLING fix from the Split atlas
debugging, but it also blocks D2Ds lazy-evaluation propagation: a
downstream consumer requesting only a sub-rect of our output cant
cascade that demand upstream because we keep reporting full-rect input
demand at every CustomPixelShaderEffect node in the chain.

Pass the queried output sub-rect through as the input demand. With
TRIVIAL_SAMPLING, our shaders read input pixel (x,y) for output pixel
(x,y) -- so the input rect needed is exactly the output rect being
requested.

Effects that need full output dimensions for math (Split Comparisons
center-of-image pivot, etc.) read them from host-injected OutputW/
OutputH cbuffer fields, not from the per-dispatch rect, so this change
is transparent to them. Verified Split pivot still lands at y=1080 of
2160 on the multi-branch reproducer that originally surfaced the bug.

Foundation step for upcoming preview-resolution scaling and per-input-
rect routing optimizations. 154/154 tests pass.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: Spruill-1

Effects/CustomPixelShaderEffect.cpp

@@ -256,12 +256,27 @@ namespace ShaderLab::Effects
         D2D1_RECT_L* inputRects,
         UINT32 inputRectCount) const
     {
-        // Return the FULL clamped output rect (not the clipped viewport).
-        // This ensures the intermediate textures match the output rect,
-        // so TEXCOORD maps 1:1 with GetDimensions() normalization.
+        // Pass the queried output sub-rect through as the input demand. With
+        // D2D1_PIXEL_OPTIONS_TRIVIAL_SAMPLING set, our shaders read input
+        // pixel (x,y) for output pixel (x,y) -- so the input rect needed is
+        // exactly the output rect being requested.
+        //
+        // Honoring the sub-rect (rather than returning m_lastOutputRect for
+        // every input) is the linchpin of D2Ds lazy-evaluation propagation:
+        // if a downstream consumer (e.g. the swap-chain present at a small
+        // canvas size) requests only a sub-rect of our output, that demand
+        // cascades upward and every upstream effect renders only what's
+        // actually needed. For preview-resolution scaling on heavy 4K HDR
+        // graphs this can reduce upstream pixel-shader work by 4x or more.
+        //
+        // Split Comparison and other effects that need the full output rect
+        // for math (pivot recentering, etc.) read the full dimensions from
+        // host-injected OutputW/OutputH cbuffer fields, not from the per-
+        // dispatch rect -- so this change is transparent to them.
+        const D2D1_RECT_L safe = outputRect ? *outputRect : m_lastOutputRect;
         for (UINT32 i = 0; i < inputRectCount; ++i)
         {
-            inputRects[i] = m_lastOutputRect;
+            inputRects[i] = safe;
         }
         return S_OK;
     }