perf: multi-channel image composite materializes O(n_channels) full-res RGBA cubes #734

Same "don't materialize the full N-fold representation" class as #731, for images.

Most multi-channel composite branches (render.py:2006, 2034, 2042, 2100, 2116) do np.stack([cmap(layers[ch]) for ch in channels], 0).sum(0) / n, holding O(n_channels) full (H,W,4) float64 buffers + the stacked (N,H,W,4) cube simultaneously (peak ~64·N·H·W bytes; GB-scale on large multiplex images). layers also retains every channel's normalized full-res float array at once.

One branch (render.py:2067) already proves the fix: a single (H,W,3) accumulator with comp_rgb += ... in a loop — O(1) full buffers. Apply the same streaming accumulate to the other branches (and fold per-channel normalization into the loop so only one channel is live).

Caveat: this touches the blending math (the documented /n_channels averaging design) — must produce byte-identical output, verified main-vs-branch, not just refactor the allocation.