wg_engine: improve stencil fill performance

Non-convex fills required a separate stencil-and-cover sequence for
each shape, even when consecutive shapes did not overlap. Repeating
these submissions reduced throughput in shape-heavy scenes.

Improve WebGPU performance by batching compatible stencil geometry
into one draw. Merge adjacent solid covers, preserve gradient cover
order, and retain boundaries for overlaps and render state changes.