Since Thunderbolt is also some sort of PCIe port, couldn't you have the GPU be internal to the Cinema Display, and transmit only the GPU code over Thunderbolt? Or are significant parts of the screen drawn by the CPU directly these days?

I suppose that would be possible, if the built in GPU has local video RAM and such. You wouldn't nearly be consuming the bandwidth compared to sending raw video to the display.

I do think it would be a little too complicated though, it's probably more likely there will be a faster Thunderbolt port before that happens ;-)

Latency is a much larger problem.

Is there higher latency between Thunderbolt connected components than "regular" PCIe?

Yes. Even with PCIe "extender" ribbon cables of a few inches, you can stumble over latency issues.