[AI] Replace brush with click-based interaction for AI object mask

[andriiryzhkov]

Simplify the object mask UX and improve selection quality.

Why remove brushing

The brush-first step was my initial idea to improve UX and speed up object selection - let the user quickly paint over the object instead of clicking multiple points. After extensive testing, the effect turned out to be the opposite: in many situations the initial selection was completely wrong (inverted mask, selecting background instead of object, or producing noisy fragmented masks).

The root cause: the brush approach generated 10-30+ dense point prompts from a single stroke, but SAM2 was designed for small amount of precise clicks. This caused several problems in practice:

• Inconsistent results: the same brush stroke on the same object could produce a good mask or a completely inverted one depending on exact mouse path and speed
• Inverted selections: SAM often selected the background instead of the brushed object, because the dense points overwhelmed the model's prompt encoder
• Speed-dependent bias: faster mouse movement produced sparser points, slower movement denser - the model response changed based on how fast the user moved
• No spatial diversity: all points clustered in the brushed area, giving SAM no information about the rest of the object

A single click on the center of an object consistently produces a better mask than a careful brush stroke. This matches Meta's SAM demo and other segment-anything tools where click-to-select is the standard interaction.

Changes

• Click to select: first click on an object creates an immediate mask
• Click +/- to refine: additional clicks add/remove regions (shift+click for background)
• First click is always foreground: shift is ignored on the first click
• Removed brush code: brush painting, brush cursor, brush points, brush size config
• Renamed brush_used to has_selection: cleaner semantics
• Added feather slider: adjustable path border/softness

[TurboGit]

Works for me, thanks!