In gegl_parallel_distribute_{range,area}(), make sure not to use
more threads than there are pixels to process (or, in the case of
gegl_parallel_distribute_area(), rows/columns to process), so that
each thread has at least some work to do.  Since thread cost is now
calculated dynamically, low thread-cost values could result in too
many threads being used, some of which would process an empty area.
This could lead to bugs, on top of being inefficient.