Why does the standard Python function os.cpu_count() often return the total number of cores of the physical server inside a Docker container running in a Kubernetes pod with CPU limit: 2? What critical problems with thread pool size in Gunicorn or Celery does this cause, and how to fix it?