I think the highest performing card you could get that meets all these is the 8600GT, which has a TDP of 47 watts (source) – though the actual power consumption is probably higher, and thus over your limit. Maybe look at the 8400GS? I think that’s the lowest-end card that still supports CUDA.
If you get one of those, and it draws too much power for the dock, you could always under-clock it as well.
Also, if it has an x16 connector, but only supports x1 data transfer…I don’t even know that it would be worth using, except for a test-bench, since the data transfer would be so slow (PCIe x1 = 250MB/sec).