I came to the conclusion that the V1 of those boards didn't mux the I2C, so all N sensors see the same transactions, and the mechanism for NAKing invalid messages will be broken if even one sensor sees it correctly. It's also why it doesn't work on V2.1 camera boards as we read unique data back from those modules (V1.3 boards I think we only read the sensor ID, which will always be 0x5647).
Seeing as they're advertising a V2 board that supposedly works, I guess they've made all except one I2C bus "unidirectional", so sensor 1 replies on behalf of all of them.
The resync interval on switching is undefined. You will certainly corrupt the frame that was being received during the switch, and I couldn't say for definite that you won't get that corrupted frame delivered (there are CRCs on each line, I can't remember about the whole frame).
Watching the youtube clip embedded in the IVmech page
I see a glitch about 42 seconds that appears to be part of two different frames, so I'd say that's a definite that you can get through partially corrupt frames. The magenta frame at about 1:10 and other points are also a classic symptom of CSI2 corruption.
Auto gain and exposure control will be quite badly broken if you're switching between sensors looking at significantly different scenes. AWB will have no knowledge of the switch either. About 56 seconds in that youube clip is a good example of that where it is having to take several frames to adjust after a switch from the desktop and hands to the office wall and posters.
There is also some latency in the GPU side, so on switching the GPIOs from userspace there is no guarantee as to exactly how many frames later the resulting frames will have switched to the new sensor.
Currently software support would have to be part of the GPU firmware, and that isn't going to happen I'm afraid. I don't have the spare time, and RPF are unlikely to see the business case for them to fund it.
There is the work ongoing to expose the CSI2 receiver via V4L2, but you then lose the ISP and will be getting raw Bayer frames, or have to use an alternate sensor with built in ISP.