You are correct, it would take a lot of effort to push the block reward down to 5. But while you noted the 25x speed increase posted earlier, that was only for thr sieve portion of the coding. According to Sunny, the sieving and fermat testing were pretty much split 50/50 for time, so that 25x speed increase could end up only being 2-4x at this time. Admittedly, you could change parameters to sieve deeper and get different results, but that would take a lot of trial and error. Unlike SHA and Scrypt where a change is virtually immediate, the finding of prime chains is more random and can take days of testing to really see results.
Yes, but only sieve part is highly efficient already not because fermat testing is not implementable on the GPU, but because GPU developer hadn't enough time to finish it, he just did sieve part first. I've understood that in fermat testing gains may be as big as in sieving if not higher.
Nevertheless, such high gains wouldn't be needed at all to kick the VPS farms out of the game. I don't have the VPS farm to testify how much profit they have per $ invested at current difficulty, but I'm sure even more moderate (than 25x) GPU efficiency gains will bring VPS farms in red.