Revision 39088

  Modify the way the packet size is calculated following a simple formula aiming at
  having all workers finishing at the same time:

                       n_j =  r_j / Sum r_i * N_left

  where 'Sum r_i' is the total instantaneous rate, 'r_j' is the instantaneous rate
  of worker j and Nleft is the number of cycles left to be done.

  The patch also changes the way the size of the calibration packet is calculated:
  a fraction is now used instead of an absolute number, which was too small in some
  cases; the fraction refers to the average number per worker and defaults to 1%; it
  can be changed with the Double_t parameter 'PROOF_PacketizerCalibFrac', e.g. 

                 proof->SetParameter("PROOF_PacketizerCalibFrac", 0.05)

  will set it to 5% .

  This patch solves the problem observed in the ALICE CAF where the variations of
  performance across machines was not correctly taken into account, and increasing
  workers was not always leading to an increase of the overall perfomance.