Render: the decoded-image cache can grow to a quarter of installed memory on every desktop OS

Squashed from #5034 onto main.

docs/target-app-parity.md

@@ -1382,6 +1382,7 @@ Moved from the ROADMAP's "Shipped so far" on 2026-10-10. Update it when a featur
   Hard-edge rendering, including the internal hard-edge frame of Art Optimized supersampled raster exports, draws a knockout element with one opaque solid fill once instead of erasing and adding it through separate layers. Analytical antialiasing keeps the existing compositor and rounding. Synthetic path/compound-hole tests compare exact pixels against the general knockout path across precision, supersampling, translucent parent groups and backdrop blending; translucent fills retain their knockout behavior.
   An image whose pixels can't be decoded, such as a PNG cut short or an image over the 2 GB decoding allowance as RGBA that can't be drawn at reduced size, is tried once: the renderer notes it and does not try to decode it again on later frames. An image refused only because the images being drawn leave no room for it is tried again on the next render. On the canvas such an image draws as a light gray box crossed from corner to corner, and the status bar says once for each image that it couldn't be decoded; exports leave it out and report it as before.
   A PNG, JPEG, GIF or WebP image that takes more than 2 GiB as RGBA (512 MiB in the web app) draws in the canvas and in raster exports at the largest size within that allowance, instead of as that crossed box on the canvas and not at all in exports. Each pixel of the result is the mean of the image's pixels it covers, and the document keeps the whole file. A PNG that isn't interlaced is read a row at a time, and other images are read whole when their decoded samples fit in the allowance. A 16-bit PNG that isn't interlaced, whose decoded samples take more than the allowance and whose RGBA fits in it, draws at its own size, read a row at a time. An image of more pixels than the allowance has bytes (on the desktop, 46,341 × 46,341 is more) still draws as the crossed box, and so does one whose decoded samples take more than the allowance, other than a PNG that isn't interlaced. Open, Place and the Links panel say "an image of W × H pixels is shown at reduced size", and say "too large to draw" of an image that is not drawn at reduced size.
+  The renderers keep the decoded images they share within a quarter of the installed memory on Windows, macOS, Linux and FreeBSD (on Linux, of the process's cgroup memory limit where that is lower, and in a 32-bit build at most a quarter of the address space), within 1 GiB where the installed memory can't be read, and within 256 MiB in the web build; while the decoded images add up to less than that budget, they are not decoded again when the view changes.
   With `--deep`, each memory row of `vectorcraft-cli perf` also gives the peak so far: the peak working set on Windows and the peak resident memory (`VmHWM`) on Linux. A new last row gives the memory after the SVG export and the save → open → save at the end of the run.
   The renderer composes the two ink planes of a CMYK document that shows transparency into screen colors on up to as many threads as it rasterizes with, instead of on one. At 2880 × 1800 on 8 threads, a page of 300 shapes at 20% to 90% opacity takes 30 ms a frame instead of 55 ms, and 50,000 paths with one at 99% opacity take 258 ms instead of 361 ms (`vectorcraft-cli bench`), with the same pixels.
   With Overprint Preview or Separations Preview on, the renderer keeps the art it prepared for the preview from one frame to the next. It prepares again the objects that changed, and all of the art when Discard White Overprint changes or, with Separations Preview, when the plates shown, the proof setup, the swatches, the Spot Colors option or the color settings change or a profile is loaded. It draws the other objects from the art, geometry and stroke outlines it kept for them.