I think the lookback is what is killing you here. The first two REs are anchored with a constant (^, \n), so they severely limit the possible starting point. The third has to look at every character in your huge file before it can decide if it's a valid starting point (if I understand how the perl RE engine works).
So in the benchmark above, the first two have to try matching about 10,000 times (since there are 10,000 valid starting points), while the last one has to consider 480,000 possible REs (since every character could be valid).
(For the interested, try looking at the output of -Mre=debug on each of these - it was quite interesting. Thanks to the kind monk on the CB that reminded me of it!)