Default to the number of CPU cores, which seems about optimal
on my laptop. Using one more saves me 2 seconds actually.
Better packing of workers improves speed significantly.
In 2 tests runs, I saw segfaulting workers despite my attempt
to work around that issue. So detect when a worker does, and re-run it.
Removed installSignalHandlers again, because I was seeing an
error "lost signal due to full pipe", which I guess was somehow caused
by using it.
Sponsored-by: Dartmouth College's Datalad project
