Put OCIO VAE Encode into the pipeline diagram, where a grey stock box used to be

The diagram still showed the input side going through an unnamed "to VAE range" step, drawn grey because at
the time nothing in this pack did it. OCIO VAE Encode does now, so the box is ours: float32, and it reports a
value that arrives outside the range the VAE was trained on, which the stock encode passes through in silence.

Both descriptions of the same picture were updated with it, not just the picture: the SVG's own <desc>, read
by screen readers, and the alt text in README, which is a separate copy and would otherwise have kept
describing a step that is no longer there.

Checked the rest of the diagram against the code rather than trusting it, since a lot moved this week:
- LogConvert's widget really is `operation` with "Linear to Log" / "Log to Linear", curve "ACEScct". Correct.
- VAE Decode is precision float32, clamp False by default. Correct.
- prores_4444 resolves to yuv444p12le, so "12-bit 4:4:4" is right.
- the 128-channel latent matches what the VAE docs record for LTX2.
- compression now reads "compression set to zip" rather than bare "zip compression", because the default
  changed to dwaa this release and the old wording could be read as describing the default.

Text fit measured with the real font metrics rather than eyeballed: all 29 labels sit inside their boxes and
inside the page.

Author: SlavaSextron

README.md

@@ -586,7 +586,7 @@ preset and log gets treated as linear: the frame comes out flat and grey.
 
 <div align="center">
 
-<img src="docs/assets/ltx25_pipeline.svg" width="880" alt="The LTX-2.5 ACEScct pipeline in three columns. INPUT: OCIO Read of a folder of EXR frames in ACEScg scene-linear, into OCIO LogConvert set to Linear to Log with the ACEScct curve, giving ACEScct codes in 0 to 1, then scaled to the VAE's own range. MODEL: LTX-2.5, a 22B transformer with a 128-channel video VAE trained alongside it, and a separate audio VAE producing the synchronised track, which bypasses colour entirely. OUTPUT: OCIO VAE Decode in float32 with no clamp, giving ACEScct codes back, into OCIO LogConvert set to Log to Linear, giving scene-linear HDR, which feeds two writes: an EXR 16f or 32f master with lossless zip compression, and a ProRes 4444 12-bit review movie carrying the audio track.">
+<img src="docs/assets/ltx25_pipeline.svg" width="880" alt="The LTX-2.5 ACEScct pipeline in three columns. INPUT: OCIO Read of a folder of EXR frames in ACEScg scene-linear, into OCIO LogConvert set to Linear to Log with the ACEScct curve, giving ACEScct codes in 0 to 1, into OCIO VAE Encode, which runs at float32 and reports any value landing outside the range the VAE was trained on. MODEL: LTX-2.5, a 22B transformer with a 128-channel video VAE trained alongside it, and a separate audio VAE producing the synchronised track, which bypasses colour entirely. OUTPUT: OCIO VAE Decode in float32 with no clamp, giving ACEScct codes back, into OCIO LogConvert set to Log to Linear, giving scene-linear HDR, which feeds two writes: an EXR 16f or 32f master with compression set to zip for a lossless master, and a ProRes 4444 12-bit review movie carrying the audio track.">
 
 </div>