Move imported geometry GPU resources into per-renderer caches

[Copilot]

Imported geometry was storing device-bound meshes and textures directly on the shared ImportedGeometry object, unlike wad rendering which caches renderables per device/viewport. That made depth-sorted index updates unsafe when the same imported geometry was rendered from multiple swapchains or devices.

• Shift imported geometry to shared CPU-side data only

  • ImportedGeometryMesh no longer derives from Mesh<T>.
  • ImportedGeometryTexture no longer owns a Texture2D.
  • ImportedGeometry now keeps only imported mesh/material/texture source data plus versioning for cache invalidation.

• Add per-device imported geometry renderer

  • Introduced ImportedGeometryRenderer to materialize:
    • vertex/index buffers
    • input layouts
    • DX textures
  • Cache entries are keyed by imported geometry / imported texture and rebuilt when the source version changes.

• Update all imported geometry render paths

  • Panel3D, item preview, and offscreen thumbnail rendering now resolve imported geometry through ImportedGeometryRenderer.
  • WadObjectRenderHelper uses renderer-provided textures/models instead of reaching into shared objects for DX state.
  • This aligns imported geometry rendering with the existing wad renderer model.

• Remove global device coupling

  • Removed the static imported-geometry device hookup from DeviceManager.
  • Imported geometry GPU state is now scoped to the renderer instance that owns the current GraphicsDevice.

• Preserve editor/compiler consumers

  • CPU-side code that reads imported geometry mesh data (room compilation, picking, room import adjustments, bounds calculations) continues to use the shared model data.
  • Added focused tests around imported geometry model/texture version invalidation.

Example of the new usage pattern:

var model = _importedGeometryRenderer.GetModel(instance.Model);
var texture = _importedGeometryRenderer.GetTexture((ImportedGeometryTexture)submesh.Value.Material.Texture);

Instead of sharing mutable DX buffers/textures through ImportedGeometry itself, each renderer now owns its own GPU copy and can depth-sort/update safely in isolation.