Fix feDisplacementMap applying scale twice

[bhaskargurram-ai]

apply_displacement_map converts scale into pixels with scale_coordinates(fe.scale(), fe.scale(), ts), and then displacement_map::apply multiplied the offset by fe.scale() a second time, so the effective displacement was scale². For scale="30" that is a ±450 px offset: almost every sample lands outside the filter region and only a few specks of the source survive, which is what #1019 (and Figma/Illustrator "distort" exports) show.

This PR:

• uses scale once (P'(x,y) = P(x + scale·(XC − ½), y + scale·(YC − ½)), with scale converted from user units to pixels as before);
• reads the displacement map with non-premultiplied color values, which the spec requires ("The calculations using the pixel values from in2 are performed using non-premultiplied color values") and which the kernel's own doc comment already asked for; Skia does the same (unpremul in sk_displacement);
• adds regression tests: a solid-color map (analytically verifiable shift of scale/2), the same map at 50% opacity (must render identically), primitiveUnits="objectBoundingBox", scale="0", and the fractal-noise reproducer from the issue.

With the issue's reproducer, the rect goes from 11 opaque pixels to 10,753 (unfiltered: 10,800), i.e. a solid interior with eroded edges, matching what Chrome/Firefox/Safari produce.

Not changed here: in is still converted to the color-interpolation-filters space like the other primitives (Chrome does this too), and objectBoundingBox still averages the bbox width/height into one scalar (Chrome only uses the horizontal scale, so there is no clear per-axis reference).

Fixes #1019