CHORUS: Learning Canonicalized 3D Human-Object Spatial Relations from Unbounded Synthesized Images