There are second order benefits, like slowing down with less acceleration, so your coffee won't spill, but compared to the infrastructure investment and privacy issues, those things are really negligible.
Without the extra information, it's a completely reasonable decision to start overtaking. And then you slow down as well, after spotting the same information. So yes - V2V would help here.
Only reasonable if 'defensive driving' is a curse word in your dictionary. Otherwise it makes sense to wait a bit and see if you can detect something too. BTW, Google reported radar working 'through' at least one car, it can look through the gap between the car and the road surface. And LiDAR can see through glass of a car in front, or just over its top (the sensor bump of the car in front down' cover much of you car's view when cars are separated by several car lengths).
Well, if you can't make a car that behaves like a good human driver, then maybe you shouldn't be making driverless cars? Otherwise, instead of reducing accidents, you might increase them (at least initially, before humans adapt). Self driving cars that can't interoperate with humans using existing infrastructure are basically useless - infrastructure lifetimes are orders of magnitude longer than technology life cycles. Re-planning cities for driverless cars now will make you regret it when the technology matures.
Also, while centralized planning might fly in current day '1984' London, I sure hope this fails at least in EU, where they claim to value privacy. Nothing about self driving cars requires either centralized reporting or even V2V - if you can detect a pedestrian behind a tree, you sure as hell can measure the speed of the car ahead.