Text-to-Scene with Large Reasoning Models