fix: revert dual queue to single GPU — light worker caused 2x regression

Root cause: render-worker and render-worker-light shared the same GPU,
causing contention. Complex TRB renders went from 17s → 36s (2x slower).

Changes:
- Thumbnails back to asset_pipeline queue (not asset_pipeline_light)
- Dispatch routing always uses asset_pipeline (no queue splitting)
- render-worker-light gated behind "multi-gpu" profile — only starts with:
  docker compose --profile multi-gpu up -d
- For single-GPU setups: all rendering is sequential on one worker

The dual queue approach is correct for multi-GPU machines where each
worker gets its own GPU. On single-GPU, serial execution is faster
than concurrent GPU contention.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

backend/app/domains/pipeline/tasks/render_thumbnail.py

@@ -72,7 +72,7 @@ def _pipeline_session(tenant_id: str | None = None):
         engine.dispose()
 
 
-@celery_app.task(bind=True, name="app.tasks.step_tasks.render_step_thumbnail", queue="asset_pipeline_light")
+@celery_app.task(bind=True, name="app.tasks.step_tasks.render_step_thumbnail", queue="asset_pipeline")
 def render_step_thumbnail(self, cad_file_id: str):
     """Render the thumbnail for a freshly-processed STEP file.
 
@@ -188,7 +188,7 @@ def render_step_thumbnail(self, cad_file_id: str):
     pl.step_done("render_step_thumbnail")
 
 
-@celery_app.task(bind=True, name="app.tasks.step_tasks.regenerate_thumbnail", queue="asset_pipeline_light")
+@celery_app.task(bind=True, name="app.tasks.step_tasks.regenerate_thumbnail", queue="asset_pipeline")
 def regenerate_thumbnail(self, cad_file_id: str, part_colors: dict):
     """Regenerate thumbnail with per-part colours."""
     pl = PipelineLogger(task_id=self.request.id)