perf: FrustumCullBoundingBox SSE replacement, 1.6x speedup

- silicon_sse: si_frustumCullBBox (0x686000) — inline V4 mat*vec3
  transforms, SSE perspective divide, 4-wide horizon buffer scan.
  Benched 1.6x (117→72 cyc/call). Installed via JMP patch (544 bytes,
  won't fit in 380-byte original).
- bench: add frustumCullBBox benchmark with mapped globals, identity
  matrices, and horizon buffer test fixture.
- Remove patch table entry for frustumCullBBox, install via detour hook
  instead (allows future A/B testing if needed).

src/silicon/silicon.zig

@@ -2089,6 +2089,7 @@ const sse = struct {
     extern fn si_vec3Dot() callconv(.naked) void;
     extern fn si_translateBoundingVol(u32, u32) callconv(TC) void;
     extern fn si_processLinkedListCollision(u32, u32, u32, u32) callconv(FC) u32;
+    extern fn si_frustumCullBBox(u32, u32, u32) callconv(FC) u32;
 };
 
 const PatchEntry = struct {
@@ -2124,6 +2125,7 @@ fn getPatchTable() []const PatchEntry {
         // vec3Dot (0x602630): removed — 0.6x regression, x87 is optimal for this ABI
         .{ .target = 0x686820, .replacement = @intFromPtr(&sse.si_translateBoundingVol), .name = "translateBoundingVol" },
         .{ .target = 0x6ABC40, .replacement = @intFromPtr(&sse.si_processLinkedListCollision), .name = "processLinkedListCollision" },
+        // frustumCullBBox: A/B tested via detour hook, not patched
     };
     return &table;
 }