I would suggest that the main point of using fewer CPUs (or threads) is to keep the system more responsive for other tasks during an analysis.
Particularly with a hyperthreaded CPU (4 physical cores, reporting 8 cores to Windows) using only as many threads as physical cores seems to yield most of the performance. Note that with more threads, reported CPU time increases due to overheads, but total elapsed time (ET) reduces only very slightly.
I hadn't seen the 'cpus_to_use' option - I picked up an older one, which didn't work properly in 2.0: