This patch solves the following issues:

1) The original design of the queue expected the list to store *all*
   the dabs (copies and postprocessing ones) that were requested after
   the latest real, non-copied dab. That caused severe problems on very
   trivial brushes (with precision <= 3), which had only 1(!) dab for
   the stroke (all the other dabs were just clones). These brushes asked
   the queue to store *all* the dabs of the stroke, which ate and fragmented
   memory.

   Such requirement was ruled by the addressing trick: we could get the
   index of the job by its seqNo in contant time. Now we broke this
   requirement and started to locate the needed job with binary search
   in O(log(N)). In our case N will seldom be higher than 1000.

2) Shared pointers also solve the problem of copying the data between
   original and running job. No no copying happens.

3) KisDabRenderingJob is now split into two classes with different
   responsibilities. KisDabRenderingJob stored the data for the rendering,
   and KisDabRenderingJobRunner (derived from QRunnable) executes it.

Ref T7052