feat: convert 3D scenery models on worker threads with deferred replies

The dashboard's web server answers one request at a time, so converting a big
.nif (glom files up to tens of MB) held up every other request, flairs included.

- dWeb: Web::Defer hands a request to another thread; the reply is sent from the
  web thread on its next poll (DeferredQueue). A client that leaves first cancels
  it and the late reply is dropped. The synchronous route API is unchanged.
- Web::Shutdown closes connections while the state their close events touch is
  still alive; the destructor no longer runs handlers during static destruction
  (stopping the dashboard aborted in ~WSClient).
- WorkerPool: priority lanes, with one thread only for urgent work (flairs,
  small models, textures), and limited background work.
- Scenery: mesh and texture routes (and the showcase's) convert on the pool;
  thread-safe memory and disk caches, one conversion per model at a time with
  waiters sharing it; zones are converted ahead onto the disk cache while viewed.
- Setting scenery_workers (0: half the cores, 2 to 4).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tests/dWebTests/CMakeLists.txt

@@ -34,6 +34,9 @@ set(DWEBTESTS_SOURCES
 	"${PROJECT_SOURCE_DIR}/dDashboardServer/ai/ModeratorPrompt.cpp"
 	"WorldViewTests.cpp"
 	"NifFileTests.cpp"
+	"DeferredReplyTests.cpp"
+	"WorkerPoolTests.cpp"
+	"${PROJECT_SOURCE_DIR}/dDashboardServer/routes/WorkerPool.cpp"
 	"${PROJECT_SOURCE_DIR}/dDashboardServer/routes/NifFile.cpp"
 	"BackupFilesTests.cpp"
 	"SecurityFixesTests.cpp"