Interestingly enough, a friend suggested that the backwards retina thing actually makes sense. The argument goes something like this:

You don't actually need an exact photo. You need to be able to quickly spot the thing you're hunting, or the threat that may be coming after you, etc... Attached to the retina is a computational layer that preprocesses images (edge detection, various other bits of compression) before sending it along the optic nerve. If you sent the pure raw data, that would mean more data that needs to be sent per "image", so more effective latency.

So the computational layer is important. Now, apparently the actual sensor cells are rather more energetically expensive than the computational layer. If you put the actual image collector in front of the computational layer, then you'd have to have a bunch more blood vessels punching through the computational layer to feed the sensor layer. That is going to leave less computational power available for compressing/preprocessing it. So you're going to have to send more or the raw data to the brain and then wait around for it to process/react to it.

So now this leaves us with putting the computational layer in front of the sensor layer. But, once we have that, since the computational layer has to output to the optic nerve, well, it's pretty much got to punch through somewhere. If it simply went around, the path would be longer, so there'd be higher latency, and that would be bad. Would leave you less time to react when that tiger is coming for you.

I don't really have any references for this, it's just something that came out a conversation with this friend, but it does seem plausible to me.

(He's not a creationist, incidentally. The conversation was effectively an argument about whether intelligence is smarter than evolution, I was noting the backwards retina as an example of some of evolution's stupidity/inability to "look-ahead". He argued that actually, once one starts to look at the actual properties a vision system needs to fulfill from an engineering perspective and look at how that relates to energy flow and data latency and so on, the "backwards retina" actually might be the right way to do things after all.)