perf: optimize best fit computation and plate optimizer

- Try all valid best fit pairs instead of only the first when qty=2,
  picking the best via IsBetterFill comparer (fixes suboptimal plate
  selection during auto-nesting)
- Pre-compute best fits across all plate sizes once via
  BestFitCache.ComputeForSizes instead of per-size GPU evaluation
- Early exit plate optimizer when all items fit (salvage < 100%)
- Trim slide offset sweep range to 50% overlap to reduce candidates
- Use actual geometry (ray-arc/ray-circle intersection) instead of
  tessellated polygons for slide distance computation — eliminates
  the massive line count from circle/arc tessellation
- Add RayArcDistance and RayCircleDistance to SpatialQuery
- Add PartGeometry.GetOffsetPerimeterEntities for non-tessellated
  perimeter extraction
- Disable GPU slide computer (slower than CPU currently)
- Remove dead SelectBestFitPair virtual method and overrides

Reduces best fit computation from 7+ minutes to ~4 seconds for a
73x25" part with 30+ holes on a 48x96 plate.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

OpenNest.Engine/BestFit/GpuDistanceComputer.cs

@@ -36,6 +36,16 @@ namespace OpenNest.Engine.BestFit
                 flatOffsets, count, directions);
         }
 
+        public double[] ComputeDistances(
+            List<Entity> stationaryEntities,
+            List<Entity> movingEntities,
+            SlideOffset[] offsets)
+        {
+            // GPU path doesn't support native entities yet — fall back to CPU.
+            var cpu = new CpuDistanceComputer();
+            return cpu.ComputeDistances(stationaryEntities, movingEntities, offsets);
+        }
+
         /// <summary>
         /// Maps a unit direction vector to a PushDirection int for the GPU interface.
         /// Left=0, Down=1, Right=2, Up=3.