I'd guess that it'd allow for thinner layers which is ultimately why you can pack in more memory.

And why it's not currently done is likely because it's hard enough to stack when everything is uniform. A small deformity in the first layer will spoil the entire chip.