Adds high-precision compositing in overlay module

[masterpiga]

The overlay module previously composited its overlay through an 8-bit Cairo path. The overlay image was rendered to a linear 8-bit buffer before blending, which posterized smooth gradients (visible banding in skies and
soft shadows); the in-module hint to "move after the output color profile" was only a workaround.

What changed

• High-precision float path (default for new instances). The overlay is now rendered as scene-referred linear float in the working RGB space and composited in floating point — no quantization, no banding. dt_dev_image() gained a want_float argument; when set, gamma (which stays the terminal module) passes the 4-channel float straight through instead of packing 8-bit ARGB. A new two-stage float resampler (anti-aliased Lanczos minification + per-pixel bicubic placement) replaces Cairo for scaling/rotation. CPU and OpenCL paths included.
• compositing algorithm selector at the top of the module:
  • high-precision (v1) — float compositing (16 B/px overlay cache).
  • legacy 8bit (v0) — the original Cairo path, reproduced faithfully (4 B/px cache) so edits made before this change render identically.
• Backward compatibility. Introspection is bumped to v2; legacy_params() migrates every existing edit to legacy 8bit, so nothing changes appearance on upgrade. Only newly added overlays default to high-precision.

8 bit (left) vs high-precision (right)

Notes

• The two modes differ in color space, not just precision (legacy blends display-encoded sRGB/BGRA bytes into the linear host buffer), so legacy needed the actual old code path, not a quantized cache.
• The per-instance overlay cache tracks its format and re-renders when the mode is toggled.
• legacy 8bit still posterizes on gradients — that is the faithfully reproduced old behavior, kept only for edit fidelity.

Co-authored with Claude.

[masterpiga]

@TurboGit I have a couple more PRs ready, all about masks (as discussed on Pxls). Should I wait to send them out or do you prefer if I wait for the release? TIA!

[TurboGit]

You can create the PR now. I'll tag them for 5.8 and review will start when I have time.

[TurboGit]

I'm wondering if we want to propose this to the user. The legacy 4 bytes/pp is kept internally for old edits but all new edits should use the 16 bytes/pp.

My question is there really a need to support 4bytes/pp for new edits?

[masterpiga]

I was being conservative, but yes, I agree that it is not very useful, happy to drop it and sinplify a bit.