Show the clamp comparison as a picture, since the table alone asks people to take it on trust

The decode comparison was three numbers in a table. The figure carries the same measurement in a form someone
can check by eye: one LTX-2.5 latent, one frame, the decode path as the only variable. A patch enlarged from
both, 11606 distinct levels inside a 0.1-wide band from the stock decode against 584790 from ours, with the
stepping visible in one and absent in the other. A map of exactly which samples the clamp removes, 0.48% above
white and 0.46% below black. A difference histogram, and a scanline through the patch where the two decodes
track each other everywhere the clamp does not bite, which is the part that says nothing else moved.

Its numbers stand on their own: ranges 0.0000 to 1.0000 for the stock path against -0.0352 to +1.0408 for
ours, over 121 frames at 1280x704.

A second figure covering the write path was left out on purpose. It labels our ProRes output "12-bit", which
is the claim this release corrected: ffmpeg encodes ProRes at 10-bit whatever the format is nominally worth.
Its other numbers are sound, and its point survives without the label, but shipping a figure that repeats a
mistake the docs just fixed is worse than shipping no figure. It can be redrawn when there is a script to
redraw it with.

Author: SlavaSextron

README.md

@@ -437,6 +437,17 @@ cannot be recovered afterwards, which is why no float container downstream resto
 counts say the same thing in the domain, a reduction of about 723x, and the file size says it a third time,
 because an entropy coder has materially less information left to encode.
 
+<div align="center">
+
+<img src="docs/assets/ltx25_float32_demo.png" width="880" alt="One LTX-2.5 latent decoded two ways. Top left, frame 40 as a float32 EXR with a box marking the patch shown enlarged. Top middle, the stock ComfyUI decode at bf16 with its clamp: 11606 distinct levels inside the 0.2 to 0.3 band, and visible stepping across the gradient. Top right, the same patch decoded at float32 with no clamp: 584790 levels in the same band and a smooth gradient. Bottom left, a map of what the clamp destroys, with red marking samples above 1.0 at 0.480 percent of the frame and blue marking samples below 0.0 at 0.460 percent. Bottom middle, a histogram of the difference between the two decodes, centred on zero and spanning about plus or minus 0.03. Bottom right, one scanline of the green channel through the patch, where the two curves track each other closely. Footer: stock ranges 0.0000 to 1.0000, float32 ranges minus 0.0352 to plus 1.0408, over 121 frames at 1280x704.">
+
+</div>
+
+The patch above is the same comparison at one exposure: **11 606 distinct levels from the stock decode against
+584 790 from ours**, inside a single 0.1-wide band. The map at bottom left is what the clamp removes, 0.48% of
+the frame above white and 0.46% below black, and the scanline at bottom right is the reassurance that nothing
+else moved: the two decodes track each other everywhere the clamp does not bite.
+
 That comparison holds two variables at once. Separating them, with precision fixed at the VAE's own bf16:
 
 | decode | range | below 0 | above 1 |