I tested what you suggested, I did some ping-pong with the blur buffer. It was a worthwhile experiment, and it confirmed what I was thinking. It produces a sharper, and more aliased blur. The reason is that when using the same buffer for input and output, once a pixel is blurred it is written in the output buffer, and thus the other pixels do their blur computations with the value of already blurred pixels. This creates a very smooth blur and helps reduce aliasing. Performances are not affected by using 2 buffers.
I must admit though, that I don't fully understand how it is working, because modern GPU are supposed to do their computations in parallel, and thus it cannot ensure that all neighboring pixels of the currently processed pixel are actually done processing. My guess is that the computing time per pixel varies each frame and thus the resulting image, though the difference is so minimal that we don't see it.
In my opinion using the same buffer for input and output produces a better looking blur, and thus dof_demo3.rar (see my previous post) will certainly be the final version of my DOF shader.