It is not obvious to me that the parabola is wrong. I've certainly seen cases where reading every 2nd or 3rd block (either stochastically, or modulus) actually does take longer than reading every block, because it defeats read-ahead. But it depends on a lot on your kernel version and your kernel settings and your file system and probably other things as well.
Well, that's an interesting point, too. Maybe we need another graph that also shows the actual runtime of a bitmap scan and a sequential scan.