Well, it should depend on the purpose. If it is really "only" about finding voxels whose time course correlates with pupil diameter then you would want to go with a regressor downsampled to the number of volumes. A condition regressor or a parametric modulator thereof (or a convolution of the pupil diameter regressor with the canonical HRF) makes sense only if you are interested in the neural correlates.
> So which is the correct way, and why?
It really depends on the particular aims, so they are not correct or incorrect as such.