[AI] Replace edge refinement with guided filter in object mask

[andriiryzhkov]

Summary

• Replace custom Scharr gradient edge refinement with darktable's built-in guided filter for mask boundary snapping
• Add configurable render resolution for higher-quality edge refinement and vectorization

Edge refinement

The previous approach computed Scharr gradients and modulated the binarization threshold near edges. The new approach uses guided_filter() from common/guided_filter.h (same filter used by blend.c for drawn mask feathering). It smooths the SAM mask in flat regions while preserving sharp transitions at image edges, producing cleaner object boundaries.

Two configurable parameters (in darktablerc):

• guided_radius (0-20, default 5): filter window size. 0 disables the filter
• guided_eps (0.001-1.0, default 0.01): edge sensitivity. smaller = sharper edges

Render resolution

The render resolution is now configurable via render_size (default 1024, minimum 1024, no upper limit). The SAM encoder always works at 1024px internally, but higher render resolution gives the guided filter and vectorizer more pixels to work with, improving edge accuracy at the cost of processing time.

Users with powerful hardware can increase this for better results:

plugins/darkroom/masks/object/render_size=2048

[TurboGit]

Works for me, thanks!