Add prepare_sim.Nthread_per_load parameter

[lgarrison]

This adds Nthread_per_load as a user-configurable parameter for prepare_sim. This is used by Numba and the tree searches as the number of threads inside each of the Nparallel_load prepare_sim processes. The default is 'auto' which will use Nthread_per_load = len(os.sched_getaffinity(0)) // Nparallel_load. When setting this manually, one should make sure that Nparallel_load * Nthread_per_load doesn't exceed the number of total CPUs!

Previously, we were using multiprocessing.cpu_count() instead of os.sched_getaffinity(0), which will return too many cores when the affinity mask is in use, e.g., on shared Slurm nodes. Neither of these will give the desired answer when containers/cgroups are being used (e.g. on Binder), so in that case, the user needs to set the correct value.

[lgarrison]

This first came up in #144.