This patch was a good idea -- and addressed a big performance problem when there are multiple components, each of which has a large number of faces.
But I found that a better fix was to use a floating point version of the closest_on_tri_to_point function to see if it was at all possible that the new triangle was closer than the old, before doing the mulitprecision version. I commited that with rBdc960a81d1bd . And after that commit, I have another idea -- using bounding box tests to eliminate whole component vs component calculations, which will cut a lot more time off. If there are still cases that would benefit from parallelizing this loop after that, we can consider doing that again in the future, but for now I am abandoning the idea of merging this patch in as is. Thanks again for the patch, though!