This answer to the question about how ISO is implemented in digital cameras seems to imply that each photosite (i.e., pixel) can have its ISO set independently. If this is true, then I would think that it is theoretically possible to take a photograph in which certain photosites are at a different ISO than others. The first part of my question is: Assuming variable ISO is possible, would it be useful? It seems to me that this might be a useful way to increase the dynamic range of the sensor, e.g., by choosing a high ISO only for regions of the image that are in the shadow. Assuming variable ISO would be useful, why hasn't it been implemented in digital cameras yet? (Or has it?)
The closest thing I know to what you're thinking of is what Fujifilm are doing with DR mode in their EXR sensors, as seen in the X-10 and X-S1) - half the pixels are deliberately underexposed by a stop (or two) and combined with the "normally" exposed pixels before the image is output. For more detail, see DPReview's X-10 review - what you're interested in here is the 6 MP DR mode, rather than the 12 MP DR mode, which is the standard "underexpose and then apply a different tone curve to the whole image" seen in many cameras these days and trades off shadow noise for increased dynamic range. The 6 MP DR mode is interesting as it (in theory) allows you to increase dynamic range while keeping shadow noise as it would normally, although of course you're paying the cost in resolution instead.
Essentially, a sensor like this that would have variable exposures for each photo site would have an image that needs to be tonemapped during the RAW conversion process. More information would have to be sent with each pixel, and this would increase the size of the transmitted data, along with the processing power required in the camera. That's a mere technical issue, and I'm sure that in another few years, that won't be a problem at all.
The biggest headache I see would be making sure the popular RAW conversion programs would support the decoding process. The resulting RAW file might need to contain 32-bit color information, and there is very limited support for operating on 32-bit color images today. For the most part, they need to be tonemapped down to 16-bit first. This isn't a process that will yield great results if done automatically with today's software.
CMOS sensors are already basically an array of sensors with different ISO, which they have to compensate for. This is what gives the plastically look on CMOS sensors, but also what attenuates blooming.
However, they actually already do make CMOS chips with multiple "ISO" for achieving a higher dynamic range, where the pixel size area is double for half of the pixels, or one out of the two green pixels are twice as sensitive as the other. The cost is more transistors per pixel, which can give trouble with noise and overall sensitivity, due to leaving less space for the photo sensors. Large pixel light-integrating cells result in lower noise (generally), which is why a 36x24mm sensor at X Mpixel is better than a 1/3 inch sensor at X MPixel - they respond better to the light to overcome the noise from all the electronics.