Don Weiman pointed out that the benchmark was slower than it should be for the 4 data line options and correctly intuited that I was checking the busy flag between nibbles. He was exactly right. I have updated the zip file and the new benchmark data is:
8 data pins + RW 431 milliseconds
4 data pins + RW 532 milliseconds
8 data pins - RW 641 milliseconds
4 data pins - RW 687 milliseconds
The standard library routine also includes an unnecessary delay between nibbles (calling pulseEnable for each nibble includes the 100 microseconds delay there).
Thanks again to Don