Use thread pool for predict

Created by: aroffringa

Finished #3 (closed) . Solutions with predict & thread pool now look good as far as I can see. On Dawn I do as excepted see a CPU usage of ~4800% (all CPUs) during predict, with a 3-direction, many thousands of component model. Speed increase is also as expected a few tens of times (previous result of >100 seemed to be too good to be true -- and thus they were. Getting a recursively called thread pool right is actually not that easy!) Good improvement for doing DDE cals.

This thread pool could be a bit more efficient: right now it decides upon entry whether it should distribute the iterations or not (which it shouldn't when no threads are available) but when one direction finishes quickly and the nr of directions are larger than the nr of threads it could happen that not all cpu cores will be used.