Fitting running times

[Jorge-C]

(@josenavas I know I already told you this a while ago but I find it
much easier to keep track of stuff if it's in the issue tracker!)

In curve_fitting there's a while loop that, even if it would always
exit in exact arithmetic, is not guaranteed to do so in floating point
math.

A much more robust (and faster) method is to take logarithms of the
data and fit that to a straight line whose slope tells us if the
algorithm is linear, quadratic...

This is very robust, and besides handling all the currently handled
cases, it would also fit negative powers (think time vs number of
processors).