I think it's to do with the lighting model. Even with opacity zero, there is "mesh" there, and I think it still interacts with Rays, unless the object itself is set not to, which would of course destroy the lighting model for something like a tree (lighting model is calculating the whole square, but we just want the leaf calcualted). So with cycles, this may cause lighting issues that aren't as realistic, especially with raytracing the shadows and calculating all the opacity maps and angles towards light.
I also think in Blender, at least, with a particle system, leafs cards culled to a mask might even be faster. Not only is the area larger on the card to calculate lighting for, but the mask has to be calculated with all the cards angles and light passing through (which also may become secondaries or something, not sure).
Additionally, a leaf card actually culled to a leaf, and processed, may provide more accurate normals for the leaf as apposed to a card, due to the new geometry of the card. Especially if you add in some roughness to the cards shape for leafs sides hanging down or up ever so slightly, etc. In TG the normals for cards let alone trees are pretty crucial for good lighting. Too large, like xfrog vanilla OBJ models, and things are very fake and flat looking. I have subdivided all xfrog models, and computed new normals, and they look a lot better in PT similar to what I like in sjefen's scenes.