On Computing Universal Plans for Partially Observable Multi-Agent Path Finding