I agree, the ISO3200 modes must do binning or something very similar. This must be done quite early in the pipeline (i.e. in sensor readout like true binning or somewhere shortly after by the Digic DSP), because it's faster than normal shooting. Trying to get the same effect on the main ARM CPU where CHDK runs would be very slow. It's possible that someone could figure out how to control the hardware that does this, but no one has done it yet.
That said, if you just shoot a full resolution image and downscale it to 640x480 on your PC, that will also smooth out a lot of the noise. I'm not sure hardware binning to that resolution would gain you much except speed.