perf(bench): introduce standardized 25-metric CAD and UI benchmark suite

- Add comprehensive performance benchmark harness in benches/performance_benchmarks.rs
  configured with harness = false in Cargo.toml (runs via cargo bench in --release,
  isolated from cargo test).
- Benchmark coverage across 25 distinct CAD & UI operations:
  * Entity Ingestion, GPU Wire Partitioning, Analytical & Wide/Tapered Arc Tessellation
  * 100x Zoom Navigation, Spatial InteractionIndex, Multi-mode Hit-Testing, OSnap Tracking
  * 2D Parametric Dogleg Constraint Solving, DXF Import/Export Throughput
  * UI Ribbon View Construction, Viewport Grid Projection, Grid Cache Evaluation
  * Themed SVG Icon Lookup Caching, Cold Extents AABB Calculation
  * Batch Entity Transformation, Draworder Depth Map Full/Patch, Delta-Undo Recording
- Expose internal calculation hooks with #[doc(hidden)] pub to prevent API leakage.
- Remove 644 lines of legacy, scattered ad-hoc #[test] #[ignore] benchmarks across src/.
- Add developer and regression documentation in benches/README.md.

src/ui/icons.rs

@@ -159,7 +159,8 @@ thread_local! {
 /// trigger a `RefCell` panic on the second `borrow_mut`, because
 /// no `borrow_mut` is held while parsing runs. (The same
 /// `thread_local!` + `RefCell` shape is used by `SEMANTIC_CACHE`.)
-fn themed_handle(bytes: &'static [u8]) -> svg::Handle {
+#[doc(hidden)]
+pub fn themed_handle(bytes: &'static [u8]) -> svg::Handle {
     let key = (bytes.as_ptr() as usize, bytes.len());
     if let Some(handle) = THEMED_CACHE.with(|cache| cache.borrow().get(&key).cloned()) {
         return handle;