Saving a 36 MP document spent most of its time packing the merged image's
RLE rows one after another on one thread (packbits over ~144 MB), and
importing it decoded each layer's channels sequentially. The merged-image
row packer now runs in parallel chunks that concatenate in order (rows are
independent; errors surface like the sequential path's), the layer gather
runs in row bands on all cores before the per-channel RLE, the layer's
channels decode side by side, and the full-surface clone per layer is
gone.
Measured on a 6000×4000 8-bit document with a photo, a paint layer and a
Levels adjustment: save 2.86 s -> 2.55-2.64 s, open's import stage
1.12 s -> 1.07 s (psd_to_document 745 -> 564 ms). The remaining save time
is dominated by the merged composite render itself (~0.7 s).
