Machine-learned uncertainty quantification is not magic: Lessons learned from emulating radiative transfer with ML