feat: GPU optimization — lighting, viewport CA, benchmark mode, 531 FPS

- Resolution: 8x8 world cells, 2x2 UI cells (UI_SCALE=4)
- GPU lighting: vertex-shader computed, light source list buffer (max 64)
  instead of O(N×R²) grid scan, O(N×S) per cell
- Viewport-aware CA: iterate only active chunks, not all 250×250
- Flat array entity/item/shadow maps instead of HashMaps
- Flat 128-entry ASCII atlas array instead of HashMap lookup
- Partial grid upload: viewport + 30-cell margin only
- Pre-allocated viewport arrays in renderer structs (zero alloc/frame)
- Skip CPU lighting for GPU modes (pass None)
- Benchmark mode: --mode benchmark with per-subsystem timing
- GpuLightSource struct, light_count in push constants
- gather_sources_in_range() for viewport-scoped source gathering

Benchmark (600 ticks, release):
  Graphics: 531 FPS (was 386, +38%), render 1013us (was 1699us, -40%)
  ASCII:    402 FPS (was 313, +28%), render 1502us (was 2346us, -36%)

All 171 tests + 14 scenarios pass.

tests/perf_corpses.rs

@@ -0,0 +1,49 @@
+use verbatim::ai::AiAction;
+use verbatim::ai::GameSession;
+
+fn setup() -> GameSession {
+    let mut s = GameSession::new_seeded(42);
+    s.init_empty();
+    s.clear_area(90, 90, 80, 80);
+    s
+}
+
+#[test]
+fn corpses_do_not_cause_lag() {
+    let mut s = setup();
+    for i in 0..20 {
+        s.perform_action(&AiAction::Spawn {
+            kind: "goblin".into(),
+            x: 100.0 + (i % 5) as f32 * 3.0,
+            y: 100.0 + (i / 5) as f32 * 3.0,
+        });
+    }
+    s.step(20);
+    let ids: Vec<u32> = s
+        .get_entities()
+        .into_iter()
+        .filter(|e| e.kind == "Goblin")
+        .map(|e| e.id)
+        .collect();
+
+    let start_before = std::time::Instant::now();
+    s.step(60);
+    let before_ms = start_before.elapsed().as_secs_f32() * 1000.0 / 60.0;
+
+    for id in &ids {
+        s.perform_action(&AiAction::DamageEntity {
+            id: *id,
+            amount: 100.0,
+        });
+    }
+    let start_after = std::time::Instant::now();
+    s.step(60);
+    let after_ms = start_after.elapsed().as_secs_f32() * 1000.0 / 60.0;
+
+    assert!(
+        after_ms < before_ms * 3.0,
+        "corpse simulation should not be dramatically slower: before={:.2}ms after={:.2}ms",
+        before_ms,
+        after_ms
+    );
+}