Changes from version 1.02:

- fixed a bug. By default, a weighted average of solutions and lower bounding planes is taken (for iterations 0,1,2,...).
  The first lower bounding in the sequence was taken to be zero. This is valid when flag zero_lower_bound is true
  (e.g. in SVMs), but may be false in general. Thus, the lower bound may have been false (although the effect would probably diminish over time). 
- fixed a bug: when avg_flag == 3, the averaging was performed not as described in the paper, but a bit differently.
- added weighted averaging in the case when kernel_max > 1.

Changes from version 1.01:

- previously, the sum of vectors was recomputed after every 5 passes, and this
  affected the stopping criterion. Changed that.

Changes from version 1.0:

- fixed a bug: if \mu != 1 then the lower bound may have been computed incorrectly
- added an option to generate a random permutation after every pass (rather than after every iteration)

