I like, and understand your thinking.
On a practical basis, I don't believe that there is any way to make it work.
1. The GPU on the video card uses a specific command set, and is designed to process only video related operations.
The GPU command set, and the video card BIOS pretty much limits the command list for the GPU.
2. Communications: While all data busses can process 2 way communications, very few video cards are set-up to transmit data back from the GPU to the CPU.
For the most part, beyond video card identification, the GPU only sends a wait signal pack to the CPU.
3. Buss translation WILL slow the process down.
Even if you can send commands to the GPU, there are the buss translations and speed issues to deal with.
If this is a PCI card, the fastest buss transfer speed will be around 33MHz.
Unless you have a really slow system (386) the CPU will spend more time transmitting and receiving the data back from the GPU than it would take to perform the task itself.
System processes that take place wholly inside the CPU run the fastest.
Any process that uses external resources slows things down measurably.
So while it would be a great idea to implement distributed computing among the various dedicated processors throughout the PC, it just isn't part of the present design structure of the data busses or cards themselves.
Chas