bug#13949: 24.4.1; `fill-paragraph' should not always put the buffer as modified

[Jaakov]

> It's less likely that the before/after `M-q' strings hash to the same
> md5 than cosmic rays reprogramming your Emacs into vi, so:

Users' texts are not completely random, but highly correlated. 
Therefore, the argument about unlikeliness (which might hold for 
equidistributed texts, for each length) does not apply.

If the above patch stays upstream, I'm going to switch from emacs to 
something else and advise others to do so. Really.

Please don't cc me any more.